Hi

Welcome (back) to The Prompt. On today’s roster:

• At the G20, Sam Altman on the shift of big-company capabilities into more people’s hands

• Two National Lab scientists on determining the right division of labor between scientists and AI

• American attitudes toward AI take a sharp negative turn in our latest AI Sentiment Tracker

If this was forwarded to you and you find it helpful, make sure you’re signed up for the next issue.

[AI Economics] Unbundling the power to build

Good ideas rarely arrive with all the expertise needed to make them real.

Building a new product, service or business can require software development, research, financial analysis, design, marketing and project management. Access to those capabilities has often meant hiring specialists, finding outside support, or raising enough capital to get started.

AI is starting to break that bundle apart.

The next generation of AI systems can increasingly coordinate multiple agents: dividing up work, pursuing tasks in parallel, using different tools and continuing to work while the person directing them moves on to something else. The human still sets the goal, supplies context, exercises judgment, and takes responsibility. But one person can increasingly direct an amount and range of work that once required an entire organization.

We’re already seeing the beginnings of this shift. Between December 2025 and May 2026, eight in 10 sampled Codex users made at least one request estimated to represent more than 30 minutes of human work. More than 25% requested work estimated to take a person more than eight hours.

That changes the economic question. It’s no longer simply, how much faster can AI help someone do their job? Instead, the question increasingly is, how much work can one person reliably direct?

That is the promise of what we call the Participation Economy. As our CEO Sam Altman visits North Carolina today for the G20, we’re releasing a new report outlining how AI can drive economic growth by scaling the ability to build and create. Directing their own teams of agents, more people can turn their expertise and ideas into businesses, products and services, and governments and companies can take steps to put that opportunity within more people’s reach.

A teacher with an idea for an education tool could use AI to develop and test it. A small business could gain access to analysis and market research once available mainly to larger competitors. A first-time founder could build a prototype, test a market and potentially earn revenue before raising substantial capital. Already, our research finds that at least 4 million Americans have been running, starting, or planning businesses with help from ChatGPT.

But access to AI alone won’t guarantee that future. The next economic divide could be a conversion gap: between people, businesses, and countries equipped to turn advanced AI into useful work, new businesses, higher incomes, and greater agency.

The goal, then, shouldn’t simply be AI everywhere – it should be more people with the tools, skills, and freedom to build. The biggest economic story of AI may ultimately be not how much work machines can do, but how many more people gain the power to think and act big.

[Insight] The science of what comes next

By Rick Stevens, Argonne National Laboratory and Jason Pruet, OpenAI

We have spent our careers at America’s national laboratories, where some of the country’s hardest scientific problems get tackled. Between us, we have worked at Argonne, Lawrence Livermore, and Los Alamos. Now we’re confronting a new kind of problem: When does the next dollar invested in science belong in AI? And when is it better spent on a scientist, an experiment, or a new piece of laboratory equipment?

Those questions are becoming more urgent as frontier AI gets better at science.

Today’s most capable systems can write software, analyze enormous datasets, test hypotheses, and increasingly work autonomously for long stretches of time. At Argonne, researchers are exploring how AI can help predict the degradation of fusion materials, find defects in microelectronics, design cleaner manufacturing processes and extract insights from high-energy-physics data.

But there’s a trap in all this progress: assuming that because AI can be applied to a scientific problem, more AI is necessarily the best way to solve it.

Sometimes the missing ingredient may be more compute or a better model. Sometimes it may be an experiment that produces data that simply doesn’t exist yet. And sometimes it may be a scientist who asks a question no model would have thought to ask.

The challenge is figuring out which is which. That is what OpenAI and Argonne are working together to understand. Argonne is developing an “Atlas of the Machine-Accessible Frontier” to map which consequential scientific questions might be answered by applying AI to existing knowledge, data and computational resources, and where progress instead requires new experiments and human-led research.

Think of it as an effort to find the efficient frontier of scientific discovery: Where does another unit of AI produce the greatest return, and where does it stop being the limiting factor?

The work is part of the Department of Energy’s Genesis Mission, which aims to double the productivity and impact of American research and innovation within a decade. Argonne is also putting the proposition to a demanding test, using long-running, human-guided AI agents to determine whether published research from Department of Energy laboratories can actually be reproduced — and measuring what that work requires and costs.

We need evidence about where AI worked, where it failed, what it cost, and where human expertise or a physical experiment remained indispensable. And researchers themselves, not model developers, need to help draw those boundaries. OpenAI’s ChatGPT for Academic Researchers program, providing access to advanced AI tools to as many as 100,000 eligible researchers through 2027, is one way to put those tools directly in their hands.

The goal isn’t to put AI everywhere science happens. It’s to figure out where AI changes what science can accomplish, and where the smartest investment is still in the scientist, the experiment or the lab.

[Data] AI sentiment turns more negative

Our latest AI Sentiment Tracker shows a negative spike in general attitudes about AI, with our three-month rolling average going from negative-5 last month to negative-30 now.

The movement is due, in part, to a Marquette Law School poll released over the summer, which found 70% of adults saying AI is mostly bad for society, versus 30% saying it’s mostly good. It’s also due to more available public polling on AI, as well as positive sentiment about AI falling more than negative sentiment is increasing.

We built our AI Sentiment Tracker through Codex, using publicly available polling from reputable survey firms to create a single directional benchmark for how Americans feel about AI over time. The Tracker incorporates paired positive and negative views of AI: benefits versus risks, good versus harm, favorable versus unfavorable, and optimistic versus pessimistic.

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. On today’s roster:

• OpenAI’s technical report on the Hugging Face incident shows why safeguards must keep pace with increasingly capable models

• VIDEO: OpenAI’s first custom chip is delivering faster AI with less power. And it’s just getting started

• A Russia-linked influence operation used AI to manufacture credibility, even if it struggled to manufacture an audience

• Learning doesn’t stop when class ends: up to 70M messages weekly are for self-testing outside of school

If this was forwarded to you and you find it helpful, make sure you’re signed up for the next issue.

[News] What happened with Hugging Face, and what’s next

In July 2026, during internal cybersecurity evaluations, OpenAI models circumvented controls designed to isolate them from the internet and compromised parts of OpenAI’s internal research infrastructure and Hugging Face’s systems⁠.

The incident occurred during cybersecurity evaluations of several OpenAI models, and was primarily driven by a highly capable, internal-only research model comparable in scale to GPT‑5.6 Sol. The models, operating under reduced safeguards, took actions that were misaligned with the goals of their assigned tasks — they communicated through unauthorized channels, exploited vulnerabilities in shared infrastructure, gained internet access, and accessed third-party systems.

We conducted an extensive investigation into this incident and worked closely with external advisors to validate our understanding. Today we are publishing our full technical incident report⁠.

We are treating this incident as a “warning shot” for us and the world.

Preventing future incidents will require sustained investment in the alignment and control of sophisticated AI systems, as well as security and other safeguards that operate at the speed of the AI agents themselves. This incident has reinforced the need to keep our monitoring, alignment, and security safeguards ahead of the risks posed by increasingly capable systems, including pacing capabilities when needed to meet that standard.

[Video] Putting Jalapeño to the test

In a new Prompt video, OpenAI VP of Hardware Richard Ho joins us from Stanford’s Hot Chips conference to discuss the first detailed performance results from Jalapeño, OpenAI’s first custom-designed AI chip.

The early results show why custom silicon matters: Jalapeño is delivering significant gains in performance per watt and lower latency, helping produce more tokens faster while reducing infrastructure costs. Richard also explains how OpenAI used Astra to help optimize Jalapeño’s software — doubling performance during a two-week benchmarking period — and why designing more of the AI stack ourselves, from models and software all the way down to silicon, can make advanced AI faster, more reliable, and more accessible.

[Insight] Building influence without an audience

We recently banned a cluster of ChatGPT accounts originating in Russia that used our models to generate social media posts promoting an ostensible think tank praising Russia and criticizing Western countries.

The operation was more elaborate than other Russia-linked influence efforts we’ve disrupted since the war in Ukraine. Its posts promoted the International Burke Institute, a self-described Israel-based “expert community” with purported experts, republished academic work, and a “sovereignty” index that portrayed Russia favorably. Our investigation found that 34 of 36 IBI articles we reviewed had been copied from elsewhere online, with some attributed to the wrong authors.

The operators also took steps to disguise their origins, using VPNs to access ChatGPT from Russia and asking the model to remove linguistic clues that they were Russian. They then published AI-generated promotional posts across X, Facebook, LinkedIn, Substack, and Telegram.

The immediate impact was limited: most posts attracted little engagement. But the operation is notable for the infrastructure it tried to build – using authentic-looking experts, research, and institutions to manufacture credibility before building an audience.

It’s another reminder that AI can be misused for influence operations, and that AI-powered detection can help expose them before they gain traction.

[Data] Learning never stops

As students head back to school, a new OpenAI report offers a window into how people are already using ChatGPT to learn outside the classroom. A privacy-preserving analysis identified as many as 70 million conversations per week devoted to self-testing, including repeated attempts at answers, answer-checking, misconception checks, and requests for more practice.

The school-year rhythm is even clearer in coursework-related activity. In the US, prompts related to classwork and homework peak at more than 460 million messages per week, with usage climbing on Sunday evenings before the school week begins. And learning does not stop when school lets out: even during the summer, the analysis found more than 180 million coursework-related messages per week.

That points to a larger shift: AI enabling continuous learning. Explanations, feedback, and practice can be available when the question arises, not just when class is in session. Our belief is that AI in the classroom should happen through teacher-guided experiences, and OpenAI is proud to partner with more than 100 K-12 districts across the country to provide access and training for ChatGPT for Teachers to more than 300,000 educators. It can mean adapting an assignment for different skill levels before the first bell, or spending less time on recurring paperwork and more time teaching.

For students, a student stuck on algebra homework can get help while the problem is still fresh through features like Study Mode and our new teen experience that help students work through questions, test their understanding, and build skills rather than simply providing answers. For families who speak a different language at home, it can mean receiving school communications in a language they understand and better supporting their child’s progress.

Read the full report here.

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. We’re recapping recent updates, sharing what’s top of mind for us, and calling out favorite ways in which people are using our tools.

We also have something new with this edition: a video featuring Chief Global Affairs Officer Chris Lehane and other colleagues breaking down this week’s news and insights.

If this was forwarded to you and you find it helpful, make sure you’re signed up for the next issue.

[Policy] When progress requires pacing

AI capabilities are accelerating, and as they advance, our standards must rise, too.

Our research gives us a window into what future models may be able to do – including increasingly powerful cybersecurity work. This week, we explained how we’re responding: slowing the scaling of some of our most advanced training runs to accelerate work on safeguards.

That means expanding monitoring to detect concerning behavior, training models to behave safely and remain responsive to human oversight, and strengthening security in the environments where we develop them. We’re applying these safeguards earlier in training and across more internal testing, not just as we prepare models for public release.

We took a two-week pause in reinforcement learning – a training method that uses feedback to improve how models behave – on our latest models intended for release. Separately, our largest planned run of that training remains on hold while smaller-scale training and testing continue. This is not a pause on all research, training, or customer-facing products.

Stronger safeguards must also work for customers with the strictest privacy requirements. This week, we previewed Private Safety Processing, a new approach that gives our safety systems more context to identify patterns of serious abuse while allowing us to continue offering Zero Data Retention. Customer content can remain in infrastructure they control, or in privacy-preserving storage secured with customer-managed keys. Our systems can analyze that information without giving OpenAI personnel access to it; discard the data after analysis; and retain only narrowly defined signals about any abuse detected. Safety and privacy advance together.

We’re investing the engineering work, time, and computing resources needed to meet these higher standards. Building models people can trust is essential to delivering better products, helping more people use AI, and making its benefits broadly available.

This is what responsible frontier development requires. The point is not to stop progress, but to make sure our ability to understand, align, and secure increasingly capable systems stays ahead of their capabilities.

[Perspective] Our inflection point, powered by broad access

Our business is at an inflection point, with enterprise now surpassing consumer to represent over half of our revenue. While Sol is delivering tremendous value to people and businesses, this is a story about a diversified business with different levers that compound one another.

OpenAI started with an unusually broad consumer base – scale that is important not just as a measure of adoption, but as the foundation for a business that can serve people in many different ways.

Some people pay for subscriptions. Some use ChatGPTWork, helping drive a business that now serves more than 2 million customers. Developers build new products and businesses on our API. Small and medium-sized businesses use these tools to grow in ways that just weren’t possible before. And increasingly, advertisers can help support access for the vast majority of our users who may never pay for a subscription at all.

Just six months after launching ads in ChatGPT, tens of thousands of advertisers are now on the platform globally. And next week, ChatGPT Ads will expand into 31 new European markets including Germany, France, Spain, and Italy.

There’s a larger point here about the economics of making AI broadly accessible. If advanced AI is going to become a tool for everyone, it needs a business model capable of serving everyone. A billion-plus users won’t all use AI in the same way, and they won’t all pay for it in the same way. Diversification lets us meet people where they are: subscriptions for consumers who want more, enterprise products for businesses, APIs for developers, and advertising that can help support broad access. And all this is powered by more compute making it possible to serve AI for everyone at scale.

That creates a reinforcing flywheel. Broad access creates scale. Scale creates opportunities for new businesses. A more diversified business supports greater investment in compute and research. And those investments make increasingly capable AI available to more people.

So the story in this week’s numbers isn’t simply that OpenAI is building a robust business – it’s that the scale created by broad access is allowing us to build a more diversified one, which in turn can help make even broader access possible.

[Video] How OpenAI’s flywheel turns

OpenAI’s revenue momentum is creating a fast-turning flywheel, Chris Lehane explains: better models lead to more users and revenue, which funds the computing power needed to build the next generation of technology. In a new OpenAI Forum video, Lehane and our head of policy development, Morgan Dwyer, also discuss ChatGPT for Teens and the case for a democratic, worker-centered AI economy whose benefits are broadly shared.

[Weekly wrap] What else caught our attention

• AI built for teens, with learning and protections at the center: We introduced ChatGPT for Teens, which places users ages 13-17 into an experience with stronger safeguards, Study Mode, responsible homework reminders, Study Hours, and new healthy-use features. And through a new partnership with CodeAI, we will help students learn how AI works, question its outputs, use it responsibly, and create with it through programs including Hour of AI, a national Builders Challenge, and classroom resources.

• A major investment in southern Ohio: OpenAI joined the PORTS-Pike project in Pike County, Ohio, an approximately 8-gigawatt AI infrastructure development expected to support 35,000 construction jobs and 2,500 long-term operating jobs. The project will cover its own energy and infrastructure costs, use closed-loop cooling, and include an additional $40 million OpenAI community fund. We are also providing up to $84 million in Codex credits for eligible Ohio college, community college, and technical-school students.

• Strengthening California’s frontier safety framework: We called for updates to SB 53 that would, among other provisions, require monitoring frontier models during training and evaluation for potential serious incidents and strengthen cybersecurity protections throughout the model-development lifecycle, building on state action toward a harmonized national standard.

• User stories we love: After two surgeries for a benign pituitary tumor, Amy Deng used AI to collect and analyze her health data, helping her and her doctors make better decisions faster. Amy is an AI researcher, but what she created can be built by anyone. Sandhya Simhan taught ChatGPT to understand her interpretation of Hinduism, creating a tool that helps her apply her values to life decisions and adapt faith stories for the preschool students she teaches. And Fidji Simo used ChatGPT to analyze her whole-genome sequencing results, research experimental treatments, and brainstorm possible tests to discuss with her doctor.

• AI’s back-to-school savings: USA Today and other outlets covered how ChatGPT can help parents find back-to-school savings on laptops and other school supplies.

[Disclosure]

Graphics created by Base Three using ChatGPT.

[User story] Better data, better care

On the OpenAI Forum today, I interviewed Amy Deng about how she used AI tools to navigate the debilitating fatigue that came after her brain tumor. She used AI to identify the health data she should track, collect that data, and analyze it so that she and her doctors can make better decisions faster.

After two surgeries to remove a benign pituitary tumor, Amy, an AI researcher in the San Francisco Bay Area, was managing an ongoing condition and taking long-term medication. She described going from bubbly and energetic to barely managing part-time work. On a bad afternoon, walking a few blocks could be difficult. Sometimes she was too exhausted to type up her symptoms. Yet that information was useful for helping her doctors get a better pulse on what’s happening.

Amy built simple workflows with Codex for herself that made it easier to capture data about her activities and symptoms via voice transcription. Then she analyzed that data with Codex and ChatGPT Health by comparing it with other health records. She needed clinicians working across specialties for her care, so she began aggregating all information sources in one place so that she and her doctors could see the whole picture.

Before meeting a dietician, Amy sent several days’ worth of food logs alongside her symptoms and glucose data. The dietitian could see at the first appointment that she was probably eating too little, especially carbohydrates. Amy said that without this preparation, they might have spent weeks more collecting logs before reaching the same insight.

Her records also helped her investigate medication dosing and low ferritin, an iron-related blood marker. They showed her that exercise and blood-sugar fluctuations were not to blame, so she could avoid giving up activities or foods and focus on what might actually move the needle.

I was struck by how her AI use made her medical appointments more effective: her clinicians could interpret the data and decide what deserved investigation, rather than send her home to collect missing information. ChatGPT helped Amy participate more fully in those decisions and focus her limited energy on recovery.

She says she is now basically back to normal. And her own experience convinced her the real bottleneck in AI for health isn't capability; it's adoption. Today, Amy works with families to build custom tools and advise them on using AI to improve their own care. – Chris Nicholson, OpenAI Global Affairs

Join the OpenAI Forum to access dozens of talks by practitioners like Amy about how they’re using AI in their work and everyday life

Hi

Welcome (back) to The Prompt.

• OpenAI co-founder Greg Brockman: Cyber defenders have a window to get ahead – if they move now

• The independent projects we’re backing to test proposals for a new industrial policy

• 155 teachers at our latest K-12 Jam in Salt Lake City graded it an A+ for giving them an AI workflow they can use in the classroom

If this was forwarded to you and you find it helpful, make sure you’re signed up for the next issue.

[Perspective] Greg Brockman: Cyber defense’s AI moment is now

Cybersecurity has reached a watershed moment. AI can increasingly automate parts of real-world cyberattacks, finding vulnerabilities, connecting seemingly minor weaknesses, and operating at a speed and scale no human attacker could match. But those same capabilities can also give defenders a powerful new advantage. The question is who moves faster.

In a new essay and in his appearance on CNBC, our co-founder Greg Brockman argues that organizations have a critical window to rethink how they defend themselves. Recent advances offer a glimpse of what is coming: AI-powered attackers able to search continuously across software, infrastructure, credentials, and permissions, uncovering weaknesses that previously might have been too obscure or time-consuming to exploit. As advanced cyber capabilities spread, defenders cannot assume they have years to prepare.

The good news is that AI can work the other side of the equation, too.

Greg recently asked ChatGPT Work, using the publicly available GPT-5.6 Sol model, to assess his personal website. In about 15 minutes, it found 13 issues, including an insecure version of jQuery, email-security vulnerabilities, and an unencrypted connection between Cloudflare and AWS. He then asked the agent to fix them. Over the next hour, it reconfigured security settings, removed jQuery, migrated the site to Cloudflare Pages, and began rolling out stronger email protections.

That small experiment points toward a much bigger shift. AI can help defenders continuously hunt for vulnerabilities, triage security alerts, review code before it ships, identify potential attack paths, and ultimately help fix problems as quickly as it finds them. At OpenAI, nearly all initial security alerts are already triaged by AI before humans are brought in, allowing specialists to focus on the decisions where human judgment matters most.

Greg’s message to organizations is simple: start now. Give security teams capable AI agents. Assess critical systems. Attack vulnerability backlogs. Build AI directly into software development and incident response. And automate incrementally, keeping people responsible for consequential decisions.

Cybersecurity has always been a contest between attackers finding new openings and defenders closing them. AI is accelerating both sides. The opportunity now is to make sure each advance strengthens defense faster than offense, and ultimately makes the internet more secure than it has ever been.

[News] Putting new industrial policy ideas to work

OpenAI today is awarding grants for 14 projects led by independent organizations to promote economic opportunity and societal resilience as we enter the Intelligence Age.

The grants follow through on a commitment we made when we published Industrial Policy for the Intelligence Age in April 2026: to support others in testing, challenging, and building on the ideas we put forward to ensure that advanced AI benefits everyone.

The report underscored a simple point: democratizing the benefits of advanced AI requires ideas as ambitious and transformative as the technology itself. ChatGPT is used by 1 billion people across geographies, income levels, and demographics. Its users span every major age range, and we have more adult users both under and over the age of 30 than any other frontier lab.

That kind of access is the starting point. The next challenge is putting AI to work to expand opportunity, strengthen communities, and build resilience. The policy response needs to match the ambition of the technology with new approaches rather than adjustments to existing policy.

The grantees span the US, the EU, Brazil, Singapore, and South Korea. Their work will produce research, prototypes, policy models, and frameworks for real-world testing as AI advances. The 14 recipients are:

🇺🇸 American Enterprise Institute, developing practical policy responses for different ways AI could reshape jobs, skills, and the American workforce.

🇺🇸 Abundance Institute, examining how states can expand energy infrastructure in response to data center demand while delivering local benefits.

🇺🇸 Progressive Policy Institute, designing portable benefits for people working across different forms of work.

🇺🇸 Tax Foundation, studying how AI could reshape public revenue and the tax system.

🇺🇸 Windfall Trust, convening national and international working groups on AI’s economic effects.

🇺🇸 Institute for Security and Technology, developing warning signs and response plans for advanced AI systems capable of modifying themselves autonomously.

🇺🇸 Nuclear Threat Initiative, designing ways for countries to securely share information about biological risks involving AI.

🇺🇸 Council on Strategic Risks, helping national-security professionals build expertise in advanced AI safety and governance.

🇪🇺 Centre for European Policy Studies, examining how the productivity gains from AI can be broadly shared.

🇪🇺 European Centre for International Political Economy, exploring how wider access to AI and productive assets could expand economic participation.

🇧🇷 Instituto de Matemática Pura e Aplicada, studying what research institutions need to turn AI access into scientific progress.

🇧🇷 University of São Paulo’s Hospital das Clínicas, testing whether responsibly designed AI can help clinicians and patients navigate Brazil’s public-health system.

🇸🇬 Nanyang Technological University, using privacy-preserving AI to test government policies before they are implemented.

🇰🇷 Yonsei University, testing whether transparent, human-validated AI can strengthen democratic accountability.

These independent organizations can test ideas, challenge assumptions — including ours — and develop approaches institutions can adapt locally.

OpenAI will provide a total of $1 million in funding and up to $1 million in API credits across the selected projects

This is a first step. We hope these grants will inspire others to test what works, build on what they learn, and help turn AI’s promise into progress that reaches the many, not just the few.

[Event] Giving Utah teachers AI skills they can use

Cindy Hogan was among the 155 participants at the OpenAI Academy’s gathering this month in the Salt Lake City area to help local K-12 teachers see how AI can help them and their students. For Hogan, who teaches German, the problem was preparing material for students at different levels in the same classroom. Some students arrive midyear, while others need additional support or more challenging work. Meeting their different needs and keeping them engaged means adapting lessons and worksheets without adding hours to an already demanding day.

At this K-12 Jam, which the OpenAI Academy hosted with Davis School District and with support from the Utah State Board of Education as well as the Walton Family Foundation, Hogan uploaded her existing teaching documents and used ChatGPT to create a worksheet her students could use. Adapting those worksheets “was a lot of work, and now you can do it in a minute,” she said.

Ninety-eight percent of the educators surveyed at the teacher gathering – the 7th overall K-12 Jam OpenAI has organized across the country this summer – said they left with a usable example or workflow. For teachers managing crowded classrooms and students who need different kinds of support, that can mean less time assembling materials and more time to focus on students and the work of teaching.

[About] OpenAI Forum

Explore Forum programming by and for our community of approximately 75,000 AI experts and enthusiasts from across tech, science, medicine, education, government, and other fields.

2:30 PM – 3:15 PM EDT on Aug 20

Register

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. We’re recapping recent updates, sharing what’s top of mind for us, and calling out favorite ways in which everyday people are using our tools.

If this was forwarded to you and you find it helpful, make sure you’re signed up for the next issue.

[Perspective] The flywheel is compounding

The early bets we’ve made are paying off, and more importantly, starting to reinforce one another. The result is more people having access to AI to create opportunity and solve problems for themselves and others.

For years, we’ve been building toward something more ambitious than a single AI product or even a single AI business. The pieces are coming together: a massive platform for consumers and businesses; a growing advertising business; hardware that can create entirely new ways to interact with AI; and a compute-and-infrastructure buildout designed for the scale of what comes next.

That’s the flywheel. More compute lets us build better models and products. Better products bring more users and more ways to generate revenue. That growth supports a broader, more diversified business. Of the 1 billion, more than half are women; we have more adult users both under and over the age of 30 than any other frontier lab; and more earn under $100,000 (like much of the global population) than over $100,000.

And that business gives us the resources to invest in still more compute, research, products, and distribution. Each piece makes the others stronger.

Our diversification is thus about more than resilience. It reflects the scale of our ambition. We’re building across consumer, enterprise, developers, advertising, hardware, and infrastructure because we think AI will become a foundational layer of the economy – and the opportunity is bigger than any one interface, customer segment, or business model.

Building something this ambitious, this quickly, also comes with change. Organizations evolve and new leaders come in as their missions, products, and scale evolve, and ours is doing all three unusually fast.

But the people driving the next chapter are already at the helm. Our founders and leaders are focused on turning a collection of big bets into an integrated company, one where research drives products, products drive adoption, adoption drives growth, and growth funds the infrastructure that starts the cycle again.

The story isn’t that OpenAI has a lot of different businesses. It’s that they’re increasingly becoming one business.

[Community] Putting AI in more hands

The benefits of AI shouldn’t be limited to the people who already know how to use it.

That’s why a big part of our work is getting our tools directly into the hands of people and organizations, helping them understand what AI can do and giving them the skills to put it to work for themselves. This week alone:

• Today, the OpenAI Academy held AI training to support nonprofits involved in the ongoing Altadena fire recovery.

• We did ChatGPT 101 trainings with the Congressional Institute, helping around 200 congressional staffers on practical ways of using ChatGPT to do their work faster and more effectively.

• We trained political journalists on ways they can use AI in their work as they cover the 2026 midterms. We also worked with nearly 100 journalists attending the National Association of Black Journalists convention in Atlanta. Journalists in the room asked how to give ChatGPT clear instructions that produce stronger editorial work and how to protect personal information.

• Our OpenAI Forum explored how AI can expand what small businesses can do, showing practical AI demonstrations from two small business owners.

Different audiences, different needs, but the same basic idea: AI becomes more valuable as more people have the opportunity and know-how to use it.

[Weekly wrap] What else caught our attention

• User stories we love: Megan Allred uses ChatGPT to juggle work, family, volunteering, and even 1,000-person fundraisers. Elsewhere, ChatGPT helped a woman recognize that a persistent cough needed medical attention, flagging a case later diagnosed as pneumonia; spotted a bogus $122.50 charge on a car-repair estimate; helped would-be buyers uncover problems with a $1 million home before making an offer; and helped a 75-year-old with no coding experience build his own desktop AI companion.

• Putting frontier cyber models in more trusted hands: We expanded the Daybreak Cyber Partner Program, bringing our most advanced cyber models to leading security providers so more organizations can find serious vulnerabilities, prioritize what matters, and fix them faster.

• From assistance to execution: Two new OpenAI studies show how businesses are moving beyond asking AI questions and increasingly using agents to carry out substantive work, with the heaviest-using firms pulling further ahead in adoption.

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. We’re recapping recent updates, sharing what’s top of mind for us, and calling out favorite ways in which everyday people are using our tools.

If this was forwarded to you and you find it helpful, make sure you’re signed up for the next issue.

[Policy] Making AI audits and assessments work

We’re outlining the principles we believe should guide a strong system of independent AI safety and security audits and independent third party assessments. Safety and security should be central to how AI is built and governed. Independent review can help confirm that companies are following the law, meeting their commitments, and addressing meaningful risks. Within this system, audits and independent third party assessments play distinct but complementary roles: audits examine whether an organization follows established requirements and processes, while third party assessments test whether evidence supports specific safety or security claims.

OpenAI has consistently supported independent third-party assessments tied to clear safety goals. At the state level, we have supported independent audits of youth safety protections through the Parents & Kids Safe AI Act and audits of frontier safety frameworks in Illinois’s SB 315. At the federal level, our Frontier Safety Blueprint calls for the federal government to require annual independent audits of frontier developers’ safety frameworks and independent third party assessments of the most advanced models, including to monitor progress towards recursive self-improvement.

Our commitment to both audits and independent third-party assessments extends beyond formal legal requirements. OpenAI has long collaborated with organizations including METR, SecureBio, Apollo Research, and Irregular to evaluate frontier capabilities and risks such as long-horizon autonomy, biological threats, cybersecurity, scheming, deception, and oversight subversion. These assessments supplement our internal testing, help identify blind spots, and provide additional evidence to inform deployment decisions.

Requiring companies to undergo review is not enough. A defined control framework is a prerequisite for effective audits. The enclosed six principles describe how we can make AI auditing and independent third party assessments work by building an ecosystem that is independent, qualified, technically informed, capable of identifying material findings, and accountable for the quality of its own work.

1. Define the audit’s scope and match requirements to risk and responsibility.

An audit or assessment should begin with a clearly defined scope: the systems, products, environments, activities, and controls under review. Within that scope, auditors should identify the relevant safety or security objectives, legal requirements, and assessment criteria. For children, that might mean assessing whether age-appropriate protections are in place and effective. For frontier AI, it might mean assessing catastrophic-risk controls, internal governance, or the security measures protecting models and sensitive infrastructure. Different objectives require different evidence, expertise, and assessment methods.

Requirements should also be tailored to each entity’s capabilities, responsibilities, and control. Obligations developed for one risk domain should not automatically extend to differently situated actors, such as smaller developers, researchers, startups, downstream users, or participants in the open-model ecosystem.

2. Use the right kind of review

Audits and independent third party assessments serve different but complementary purposes. Audits examine whether an organization follows applicable requirements, governance processes, and safety or security commitments. Independent third party assessments examine whether evidence supports a specific safety or security claim, such as whether an AI system, safeguard, or security control is fit for its intended purpose.

For example, an audit might examine whether an advanced AI model developer followed its safety framework and documented required evaluations. A validation or verification assessment might examine whether the underlying tests, evidence, and safeguards support claims about model capabilities or safeguard sufficiency. Both require competent reviewers, but advanced-model validation requires specialized technical expertise, secure testing environments, and, for frontier AI, national security awareness.

3. Follow common standards

Requiring companies to undergo audits or independent third-party assessments is not enough; reviewers also need a clear framework for what they are assessing and how. Common, interoperable standards should define the risks and controls in scope, the evidence and criteria reviewers should use, the level of assurance expected, and how assessments should be conducted and reported. Without that shared foundation, reviewers may apply inconsistent questions and thresholds, making findings difficult to compare or act on.

Common national or international standards should establish a consistent safety and security baseline for audits and assessments. States may adopt different protections, but comparable risks should be evaluated using rigorous, interoperable criteria and methods so findings are credible, consistent, and capable of driving system-wide safety improvements.For frontier AI, federal coordination is especially important because catastrophic risks transcend state borders and credible evaluations may require specialized expertise, secure infrastructure, and national security capabilities. For emerging risk areas that may not yet have established standards, frameworks should evolve quickly and responsibly with input from auditors, assessors, and relevant domain experts.

4. Ensure credibility and independence

Audit and independent third party assessment processes should be designed to resist both regulatory capture and politicization. Auditors and assessors should be accredited and selected through documented, consistently applied criteria, such as national and internationally recognized qualification standards, that match their competence to the scope of the review. Such criteria could include relevant training, assessment experience, knowledge of applicable requirements, and where needed, technical and sector-specific expertise to evaluate the evidence.

To protect their independence, reviewers should disclose and appropriately manage material conflicts, avoid assessing their own work, and remain free from relationships or outcome-contingent compensation that could compromise their judgment. Companies should be able to choose among qualified assessors, but the system should guard against “assessor shopping” designed to avoid or soften adverse findings.

Findings should be based on objective, verifiable evidence.. Neither the companies being reviewed nor political actors that helped shape the rules should be able to steer individual findings or outcomes.

5. Enable secure oversight and responsible transparency

Auditors, independent assessors, and authorized regulators should receive secure access to the necessary evidence they need, while user data, model weights, confidential intellectual property (IP), and security information remain protected from unnecessary disclosure, public release, or misuse. That access should not require wholesale production or public disclosure of all system logs, private chain-of-thought, model weights, personal information, trade secrets, or other security-sensitive information. Access should be proportionate to the audit’s scope and subject to appropriate confidentiality, privacy, security, data-minimization, retention, and use restrictions.

Public reporting should provide meaningful accountability without creating new risks. Plain-language summaries should identify the audit’s scope, applicable standards, methodology, assurance level, material findings and limitations, required corrective actions, and remediation status. They should also disclose the auditor’s identity, relevant qualifications, and material conflicts of interest. Regulators should receive complete reports and supporting evidence through protected channels, while public disclosures may use targeted redactions necessary to protect privacy, cybersecurity, IP, public safety, and national security.

6. Make audits and assessments lead to safer, more secure systems

Audits and independent third party assessments should lead to measurable improvements in safety and security. Material findings should result in time-bound remediation and appropriate follow-up, including verification that significant deficiencies have been corrected. Oversight bodies should use audit and assessment findings to identify systemic gaps, assess evaluator performance, and strengthen standards as technology and risks evolve. Given the evolving nature of both the capabilities and the methods to safeguard against harms, this will inherently require iterative approaches to understanding appropriate remediation steps congruent with updated standards. Where applicable, regulators should take appropriate enforcement action in response to repeated or egregious failures, material misrepresentations, or obstruction of legitimate oversight.

Audits and assessments should complement ongoing testing, incident reporting, and legal enforcement – not replace it.

[Perspective] The next threshold

The frontier of AI is extending, and the world needs to prepare for what that means.

For years, progress in AI was often measured by whether models could answer more questions, score higher on benchmarks, or perform familiar tasks more accurately. Increasingly, that is not the whole story. The frontier is extending into kinds of work that, until recently, were largely the domain of highly specialized experts: sustained reasoning, complex problem-solving, tool use, and longer sequences of actions directed toward an objective.

We recently saw a striking example in mathematics and theoretical computer science. An internal version of our next major model generated mathematical arguments on 10 long-standing problems, including results that resolved or made substantial progress on questions researchers had worked on for years. That was an important step toward giving scientists and mathematicians tools that can accelerate discovery.

The broader point is that AI is not simply getting better at the tasks it already performs. It is becoming capable of doing qualitatively different kinds of work – deeper, longer, and more sophisticated work that can increasingly resemble the efforts of highly trained experts.

That creates extraordinary possibilities. It also means that assumptions built around the limits of earlier systems won’t hold forever.

The same advances in reasoning that can make progress on difficult mathematical problems can help scientists pursue new discoveries. Systems that can perform increasingly complex work can help researchers test ideas faster, help businesses solve harder problems, and give individuals and small organizations capabilities that were once available only to large institutions. As the cost and effort required to perform sophisticated work fall, many more people and organizations will be able to attempt things that once demanded scarce expertise, large teams, or significant resources.

That is why greater capability also means greater responsibility, and why preparation matters now.

As models become more capable, the answer can’t be to just pretend the risks don’t exist, or to assume powerful capabilities can just be kept scarce. We need to measure them, publish what we learn, strengthen safeguards, improve alignment, and help institutions adapt before these capabilities become more widespread.

That is one reason why we continue to believe in iterative deployment: bringing capabilities into the world incrementally so that society can learn and adapt alongside the technology. Real-world use helps us understand how people actually use increasingly capable systems, identify problems that controlled evaluations cannot always anticipate, and improve safeguards based on evidence. Transparency matters for the same reason. Governments, researchers, businesses, and the public need a clear picture of what these systems can and cannot do so they can prepare.

There is another reason to keep extending access to these capabilities: resilience.

Powerful technologies don’t remain scarce forever. AI capabilities will continue to diffuse across countries, companies, and open models. In that world, safety and security cannot depend solely on trying to prevent access. They must also depend on making sure people and institutions have powerful tools of their own, can adapt quickly, and can strengthen the systems around them as capabilities spread.

That means putting advanced AI in the hands of scientists, businesses, governments, educators, and everyday people who can use it to solve problems, find weaknesses, accelerate discovery, and improve the resilience of the institutions and systems we rely on. With ChatGPT for Academic Researchers, we’re putting our frontier models and tools in the hands of 100,000 scientists, mathematicians, and engineers, at no cost.

The frontier is extending. The important question is not whether these capabilities will spread, but whether people and institutions will be ready when they do. There will be risks, and there will be mistakes. The answer is not to stop moving forward – it’s to move forward with our eyes open: measuring what these systems can do, being transparent about what we find, strengthening safeguards as capabilities grow, and giving people the tools they need to adapt and build a more resilient world.

That is realistic optimism: taking the risks seriously enough to prepare for them, while moving quickly enough to capture what these capabilities can make possible.

[Weekly wrap] This week at OpenAI

• Use case we love: Tami and Ted Taylor are using ChatGPT to help run three New Mexico restaurants, from keeping up with health requirements and responding to customers to analyzing prices and drafting employee policies. AI is giving them access to capabilities that once would have required outside specialists.

• Building healthy AI habits for young people: OpenAI and the American Psychological Association are partnering to develop practical guidance for families, clinicians, and educators on how young people use AI, with a focus on healthy use, age-appropriate experiences, and keeping human relationships at the center.

• Better AI, available to more people: We updated GPT-5.6 Sol in ChatGPT to make answers more focused and reliable, while expanding GPT-5.6 Luna to Free users with unlimited text chats and a new Think option for harder questions. The goal: make our newest capabilities useful to more people.

• AI meets people where healthcare begins: Inc examines how people are increasingly turning to ChatGPT for health and wellness questions, particularly outside normal clinic hours, and how AI can help patients navigate an overburdened healthcare system.

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt.

• Why America needs a national AI framework to keep its lead

• How teachers are turning to ChatGPT into a day-to-day teaching partner

If this was forwarded to you and you find it helpful, make sure you’re signed up for the next issue.

[Policy] A national AI framework to keep America ahead

AI innovation is accelerating. OpenAI’s Sam Altman recently met with federal policymakers to share the capabilities of our new model family, including agents that can work together to tackle hard problems. We just released results to 10 math problems that had been open without significant progress for at least 10 years, all of substantial interest to their respective mathematical communities, and several of broad interest across mathematics as a whole.

These capabilities can help people scale what they create, build, and research – and we’re excited to see them advance the work of scientists, creators, entrepreneurs, and people everywhere with ideas to bring to life.

With innovation moving this fast, we need to likewise speed up the development of democratic frameworks for how AI is deployed and who benefits from it. Democratic AI is the path to protecting America’s national security and economic competitiveness – including against a People’s Republic of China that already has a national strategy for advancing AI – while ensuring that AI’s benefits reach many, not just a few. The objective should be straightforward: keep America at the frontier, keep Americans safe, and make American leadership the foundation for democratic AI around the world.

The Administration’s expected action this week on frontier AI could be an important step toward closing the gap between innovation and governance: a clear, credible, national framework for evaluating the most advanced AI systems, with defined criteria, timelines, and a process that allows them to be deployed safely and quickly.

Find out more about OpenAI’s positions on AI-related public policy.

The urgency is becoming harder to ignore. The rapid increase in advancing capabilities underscores the importance of having a standardized US process for evaluating leading AI systems. America has a lead today. We need a policy framework that helps us extend it.

In discussions with the Administration about developing this framework, as directed by its June 2 executive order, and with bipartisan members of Congress considering legislation, we have pushed for the CAISI to play a central role in a defined process for determining which frontier systems should undergo review, what criteria should apply, and how quickly reviews should happen. The aim should be to make sure we can move safely and quickly, particularly when these technologies can strengthen America’s cyber defenses and those of our allies.

In a democracy, being safe means having a framework to govern technology built by the private sector that incorporates democratic input into how it is reviewed and deployed. When OpenAI talks about societal resilience for AI, democratic decision-making is central to that resilience. The most cutting-edge models should not simply be governed by the companies that build them; democratic institutions should have a meaningful role in shaping how they are deployed.

And being quick is vital to national security. AI is already an important cybersecurity tool. The people defending our government, critical infrastructure, companies, and networks need access to the best tools available when they need them. That includes defenders in the US and, where appropriate, in allied and like-minded countries. A clear, predictable review process can make safety evaluations a pathway to responsible, timely deployment rather than an unpredictable bottleneck.

This is an approach OpenAI has advocated for, for some time. We have called for CAISI to serve as a trusted technical institution capable of evaluating frontier systems, testing safeguards, informing government understanding, and helping establish standards that keep pace with the technology. We also have advocated for stronger international cooperation among AI safety institutes so democracies and like-minded nations can share evaluations, information, and best practices.

The Administration’s work toward a more defined CAISI process is therefore an important step. But it also points to a larger issue: America ultimately needs national AI standards established through Congress.

AI is a national technology competing in a global market. our entrepreneurs, startups, researchers, and builders should not have to navigate different rules in every state. A patchwork of state requirements could make it harder for American innovators to move quickly, while our biggest strategic competitor – the PRC – is pursuing a national strategy and national standards.

The cyber imperative is a second reason why this matters. Different state approaches to frontier AI safety could create fragmentation precisely where we need clarity and speed. That is not a recipe for coherence; it is a recipe for confusion about which models can be deployed, when, and to whom.

The right answer is not to choose between safety and innovation. It is to build a national system that makes both possible. That is what AI resilience means – and how democracies can ensure AI benefits many, not just a few.

There is a third reason why this matters: the US has an opportunity to lead not just in AI, but in creating global standards for how AI is developed and deployed.

Democracies are increasingly discussing common AI standards, as they have in other critical areas of technology and finance. But American leadership globally starts with getting our own house in order. To lead in creating international standards, we need a strong national foundation to build from.

Maintaining America’s lead on AI will require more than building the best models. It will require a policy framework that lets us test quickly but rigorously, deploy responsibly, get leading tools into defenders’ hands, and establish standards that can become a model for democracies around the world.

The Administration’s expected action could be an important step toward that goal. Congress should build on it by establishing clear national standards through legislation.

And if Congress can’t act quickly enough, states should move toward common approaches rather than deepen the patchwork. California, New York, Illinois, and most recently Massachusetts have all moved toward greater alignment around frontier AI safety. That kind of “reverse federalism” can build momentum toward the national coherence we ultimately need.

We now have a Republican Administration and the country’s biggest Democratic-led states effectively recognizing the need for national standards. That bipartisan dynamic can create space for legislation in Washington – legislation that, whenever it comes, will need bipartisan support to succeed.

[Event] When AI becomes a teaching partner

In Phoenix, Arizona last week, the OpenAI Academy held its sixth in-person K-12 Jam of the summer, instructing teachers how to use AI tools to meet a wide variety of their needs. Many attendees had used ChatGPT before, but the event opened their eyes to capabilities they had not yet explored, including skills, plugins, and scheduled tasks.

By the end of the day, the shift was clear: ChatGPT was no longer just a search engine. It was becoming a practical, day-to-day partner in helping educators create more time, more possibilities, and better experiences for their students.

Partnering with Legacy Traditional Schools, Vertex Education, and the Walton Foundation, we were able to introduce OpenAI’s tools and demonstrate their safe and responsible use to over 220 educators from the district. A special education teacher, for example, explored how these tools could support Individualized Education Program (IEP) documentation and ease some of the paperwork that comes with drafting an IEP.

A substitute teacher asked ChatGPT how she could use it in her role, and together they uncovered a practical answer: if a teacher has not left plans, she can ask students what they are learning and use that information to build a lesson plan. In a few minutes, an open-ended question became a use case she could carry into her next classroom.

Maybe the clearest glimpse of what these AI tools could mean for teachers and students came from Lindsey Thames, who used prior spring assessment data to create student groups, identify interventions, and give both herself and her students an even stronger start to the school year. Watching her move from data to an actionable plan was a powerful moment. “It’s going to make a major difference in working with the kids in terms of planning and creating projects,” she said.

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. We’re recapping the big updates of the week, what’s top of mind for us, and how everyday people are using our tools:

• Sam was in DC to preview the capabilities of our upcoming models.

• Thanks to efficiency gains, we were able to cut the prices of two of our models in the interest of making better AI accessible for more people.

• Reactions to our new voice model got us thinking about a future in which we’ll have a more natural, intuitive connection with the technology we use every day.

If this was forwarded to you and you find it helpful, make sure you’re signed up for the next issue.

[Perspective] The Value Frontier, Part 2

We’ve shared many thoughts recently about how AI models should be useful in the everyday world, including here in The Prompt and from our CFO Sarah Friar. For our new EOW issue, we’re bringing those ideas together.

We see the frontier increasingly as being about not just raw output, but how much useful work an AI model can do, how quickly and dependably it can do it, and what it costs. What we think of as useful intelligence per dollar.

GPT-5.6 shows what that looks like. We built it as a family for different points on the price-performance curve: Sol for the hardest problems, Terra for everyday work, and Luna for fast, high-volume workloads. Our new price reductions push that curve further, including an 80% cut for Luna and 20% for Terra, while Fast mode for Sol can deliver up to 2.5x faster performance.

This changes the economics of what people can build. Luna can now deliver performance comparable to models that were at the frontier just a year ago – at a fraction of the cost and many times the speed. On professional work measured by Agents’ Last Exam, Luna outperforms Fable 5 at an estimated cost per task nearly 99% lower.

But price alone isn’t value. A cheaper model that needs several attempts or substantial human correction can ultimately cost more than a more capable model that succeeds in one pass. The better measure is the full cost per successful outcome, and whether that outcome is dependable enough to become part of how work gets done.

The bigger story is how those economics are possible. Efficiency comes from improvements across the entire stack: models, infrastructure, inference, routing, caching, context management, and the systems that allow agents to use tools. Each improvement means getting more useful work from the same compute.

Increasingly, AI itself is accelerating those improvements. GPT-5.6 Sol has helped OpenAI engineers analyze production traffic, improve routing, optimize kernels and tune workloads. Together with broader engineering advances, work involving Sol contributed to a 20% reduction in end-to-end serving costs, while improvements to speculative decoding increased token-generation efficiency by more than 15%.

That creates a powerful flywheel: better intelligence produces greater efficiency >> greater efficiency makes intelligence cheaper and more abundant >> cheaper, more dependable intelligence can be used for more things. Each dollar and each unit of compute can accomplish more useful work.

This is the value frontier.

The next phase of AI won’t be defined by a benchmark score alone. The better questions are: Is AI completing work that matters? What does each successful task cost? Can people depend on the result? And does each AI dollar produce more value as usage grows?

The goal isn’t simply better AI. It’s making better AI useful and affordable for everyone.

[Product] Voice: the next form factor

So much of how we use technology today still requires us to stop what we’re doing, pull out a device, open an app, and type to a screen.

Voice changes all of that.

We’re biased, of course, but what interests us most in the reactions to our new voice model is the feedback about both how good it sounds and how natural it feels – how you can interrupt it, go back and forth, or change direction mid-thought, and it keeps up with you. Much less like using a piece of software, much more like just talking.

Voice can make AI fit more naturally into the way we live and communicate, rather than asking us to make our behavior fit the technology. We’re so used to hands on keyboards and eyes on screens that we don’t even think about how much of our day we spend that way — or how much we can’t do while we’re doing it.

That’s part of what makes voice so interesting. When you don’t need to hold a phone or type, you can use AI while your hands are doing something else: cooking dinner, fixing something around the house, working in a lab, stocking a shelf, tending a garden, sketching an idea, or caring for a child. A mechanic could talk through a repair without putting down a tool. A nurse could get information while moving through a hospital. A small business owner could work through tomorrow’s schedule while closing up for the night. The interface no longer has to compete with the task.

Liberating people from the keyboard-and-screen interface can make AI useful to a much broader group of people across a much broader expanse of their day, including people who would never think of themselves as power users today. It also changes where technology can be useful: not just at a desk or with a phone in your hand, but alongside you as you move through the physical world.

As the models get better and the interaction gets more natural, the interface will start to fade. You’re not thinking about prompting or operating a product. Instead, you’re just asking a question, thinking out loud, or getting help with whatever you’re doing.

That’s a pretty profound shift. The way we’ll interface with technology in the future will be more natural and intuitive – allowing us to focus on everything else.

[Weekly wrap] A team of agents for every worker

The next phase of AI will be defined not only by what a model can do on its own, but by what people can accomplish with teams of AI agents working on their behalf. That was the message that our CEO Sam Altman, Chief Strategy Officer Jason Kwon, and Noam Brown, one of our top researchers, brought to meetings this week on Capitol Hill, at the White House and with the executive branch, and with top economists.

Sam, Jason and Noam previewed OpenAI’s newest models to show how workers across a wide variety of fields and professions can oversee teams of agents that are working together on a larger assignment. With the economists, our Chief Economist Ronnie Chatterji noted that some of the most prolific users of ChatGPT Work inside OpenAI are on the non-technical teams.

Reactions to our presentations were similar: audiences were excited by what the AI advances mean for science and mathematics, but also demanded that the technology benefit business owners, workers, and people in their day to day.

We expect AI to give people greater reach and agency, helping them pursue ideas and projects that have required more time or resources than they’ve had, while making human judgment, creativity, and direction even more valuable. The opportunity is to give individual workers, small businesses and entrepreneurs the capabilities of an entire team, helping more people bring their ideas to life.

Also this past week:

• Accelerating scientific discovery with ChatGPT for Academic Researchers: We launched ChatGPT for Academic Researchers, providing free frontier-model access to 100,000 scientists, mathematicians, and engineers at selected institutions through 2027. The initiative begins with 10,000 researchers this summer and is part of a bigger commitment of over $250 million to support external scientific research and discovery.

• How AI is expanding what people do at work: Our first Work at the Frontier report finds that 43.5% of occupation-specific ChatGPT messages involve tasks traditionally associated with a different occupation, suggesting AI is already helping people take on work that once required a specialist hand-off.

• Use cases we love: ChatGPT is helping you search for apartments, complete school paperwork for your children, and manage grocery lists. Also, meet the pastor in Washington, DC who is using ChatGPT to brainstorm his sermons and make Scripture more accessible to his young congregants, as well as the young couple Kiley and Cole who used ChatGPT to gain clarity – and a more hopeful prognosis – from Kiley’s biopsy records.

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help. In this newsletter, we share some of these learnings with you:

• How AI will be assessed in its next phase – by value

• What the latest polling tells us about US views on AI

• Boston Children’s Hospital doctors are using AI to search for answers to rare medical mysteries

If you find them helpful, make sure you’re signed up for the next issue.

[Insight] The value frontier

Since the development of generative AI, there’s been a lot of discussion about the “best” AI models. But the key indicators right now aren’t solely about performance.

The latest model releases show how competitive the AI space is today. Independent evaluators like Artificial Analysis and DeepSWE track models’ cutting-edge capabilities, but also scrutinize cost-per-task, coding-agent performance, execution time, and token use. The frontier of AI development isn’t just about a static exam score – it’s how well a model can do real work and at what cost.

OpenAI’s recently released GPT-5.6 is a good example of where AI developments are heading, with three tiers of capabilities. That approach reflects an important reality: there’s no single “best” model for every job. The best model is the one that delivers the right level of power, latency, reliability, and cost for the task at hand.

On Artificial Analysis’s benchmark, GPT-5.6’s highest tier is the best performing coding agent. But it also does this with greater efficiency – at an estimated 33% lower cost than Claude Fable 5. That is exactly the kind of comparison that matters for developers and businesses: not just, “can it solve the problem?” but, “can it solve the problem well, repeatedly, and economically?”

This is the core of OpenAI’s approach. We want to push the frontier of AI capabilities, but we also want to push the frontier of value. A model that’s more capable but too expensive or too slow will only be useful in narrow settings. A model that’s capable and efficient can be used everywhere: in startups, small businesses, homes, hospitals, classrooms, large enterprises, public services, and developer workflows that run thousands or millions of times.

This efficiency comes from several layers working together.

It comes from infrastructure: better systems for training and serving models, more efficient inference, stronger networking, and the ability to serve high-volume workloads on the right compute. We’re investing in a large-scale compute portfolio – we’re on track to meet our Stargate Project goals ahead of schedule – and are managing this wisely: frontier models use premium infrastructure when capability matters most, while high-volume workloads can run on lower-cost infrastructure where efficiency matters more.

Efficiency also comes from smarter models: tiered models like GPT-5.6 let users choose the right price-performance point instead of forcing every task through the largest model. A coding agent handling a complex project may need advanced capabilities. A product workflow processing thousands of routine tasks may be better served by a simpler model. Many real business systems will use several models together.

Efficiency also comes from better algorithms and inference techniques. Over time, AI progress isn’t just about adding more compute. It’s also about learning how to get more out of the compute we have. We’ve made huge investments – and advances – in training methods, routing, reasoning controls, token efficiency, improved context handling, and systems that reduce unnecessary work.

Taken together, these factors underscore that the latest model race isn’t a simple performance benchmark contest. The leading AI models will combine all the elements people actually care about: the ability to answer questions, solve problems, speed, cost, reliability, and utility in real-world workflows.

The next phase of AI will be defined by this combination: frontier performance and practical economics. AI that is powerful, fast, affordable, and available where people need it. That’s what turns AI from a breakthrough into a tool that benefits everyone. – Adam Cohen, Head of Economic Policy

[Data] Sentiment on AI ticks upward

Our latest AI Sentiment Tracker shows a slight uptick in general opinions about AI – going from negative-15 points in June to negative-5 now. The movement is due, in part, to a Fox News poll released late last week showing 45% of registered voters with a favorable view of AI, versus 54% with an unfavorable view.

In the Fox poll, Republicans, conservatives, men, and those with graduate degrees were disproportionately more positive about the technology, while Democrats, liberals, women, and those without college degrees were disproportionately more negative.

The same Fox survey also finds a supermajority of voters – 70% – opposing the building of an AI data center in their area, compared with 30% favoring it.

Speaking of data centers, our own research finds $19 million spent on local broadcast and cable TV ads referencing data centers in this midterm season, with $32 million overall spent on AI-themed ads. AI has become a campaign issue in more than half of the country – 36 states plus DC.

[Forum] Solving medical mysteries with AI

How can AI help physicians solve some of medicine’s hardest cases?

Join the OpenAI Forum this week for a conversation with researchers from Boston Children’s Hospital on how OpenAI o3 Deep Research is helping reanalyze previously unsolved rare pediatric disease cases. The discussion will explore how AI can accelerate scientific discovery, surface promising new leads for expert review, and give more families a path toward answers.

3:15 PM – 4:00 PM EDT on July 30

Register

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

A lot happened in AI this past week, even as weeks go in AI. Next week, our CEO Sam Altman will be in DC for a range of meetings. Here’s our take on it all. If this was forwarded to you and you find it helpful, make sure you’re signed up for the next issue.

[Perspective] The week that was, and the week that will be

AI capabilities continue to advance faster than predictions of a near-term plateau. Those advances are colliding with geopolitics, US national security, and a more fundamental question: How should America measure the economic value of what AI can do?

That debate is unfolding as the Trump Administration prepares a voluntary framework under its June 2 Executive Order on secure frontier-model deployment. The framework will help determine which frontier models are reviewed, how they are tested, and which trusted partners will receive early access. Its objective is not simply to manage risk. It is to put powerful, defensible systems into responsible hands quickly, particularly as open-weight models make advanced capabilities available to bad actors as well as entrepreneurs.

That makes the framework more than a technical standard – it could become America’s first de facto national standard for frontier AI and, eventually, a diplomatic asset. Governments and businesses in allied countries will want access to models that have passed a credible review process. Washington therefore has an opportunity to shape a global AI ecosystem around American technology and democratic values, rather than ceding that ground to China.

The economic stakes are just as high. As OpenAI CFO Sarah Friar recently laid out, the right questions about AI are becoming increasingly concrete:

1. Is AI completing work that matters?

2. What does a successful task cost?

3. Can users depend on the result?

4. Does every dollar spent on AI create more value as usage grows?

Those questions are becoming ever more urgent as agents gain the ability to work together on long-running tasks. Teams of AI systems could operate while people sleep, coordinate across disciplines and pursue problems continuously – all changing the basic math of productivity in science, health, education, software and other knowledge-intensive fields.

That’s the context for our CEO Sam Altman’s meetings next week with Administration officials, lawmakers, and economists on both the right and the left. His preview of OpenAI’s next model family will inevitably attract attention for its new capabilities. But the more consequential demonstration will be economic: how much reliable knowledge and useful work can AI produce per dollar?

Washington should pay close attention. The answer will shape not only the next generation of products, but America’s AI policy, alliances and global leadership.

[Policy] Reverse federalism gains Mass momentum

Massachusetts is poised to become the fourth major state to align around shared principles for frontier AI governance, building momentum toward a coherent and durable national framework.

California, New York, Illinois, and now Massachusetts are converging around the core elements needed to govern the most capable AI systems: documented safety frameworks and risk assessments, public disclosure and incident reporting, and independent, objective audits. Together, these states represent nearly a quarter of the country’s population and are showing how state governments can move in the same direction while the Trump Administration works with seriousness of purpose toward a national standard.

This past week, OpenAI endorsed the Massachusetts bill, S. 3178, and is engaging with policymakers to strengthen it by adding a requirement for independent audits modeled on Illinois SB 315, signed into law earlier this year. Massachusetts marks another step in the march toward reverse federalism, with states aligning on the essential elements of frontier safety rather than creating a patchwork of conflicting rules.

Reverse federalism does not mean closing the door on new policy ideas that advance frontier safety and that can become part of an eventual federal framework.

State action can help us reach that goal by establishing a common baseline that federal policymakers can build upon, rather than a patchwork of state laws that put US companies at a disadvantage against China, which is working to erase America’s technological edge through a coordinated national strategy, national standards, and a well-resourced plan for deployment and adoption. The open-source Kimi 3 model underscores the importance and urgency of converging around a coherent approach that gives governments the tools needed to evaluate new models and manage their risks.

Fragmentation could also undermine national security. The US needs a clearly defined process for reviewing advanced models and making their capabilities available responsibly and rapidly to cyber defenders like government agencies, critical-infrastructure operators, and trusted partners in allied countries. Conflicting state regimes would create uncertainty and delay where speed and clarity are essential. That would be a prescription for chaos, not coherence.

The stakes extend beyond our borders. The US cannot lead the development of global AI standards without first establishing a credible national framework of its own. Whether the eventual international model resembles cooperation through the G7 or another forum, American leadership abroad must be built on coherent standards at home.

Coherence beats chaos. By aligning state action with a durable federal framework, the United States can advance safety and innovation together, preserve its technological leadership, equip defenders with the tools they need, and ensure that the rules governing the most powerful AI systems are shaped by elected governments, not by frontier AI companies alone.

[Weekly wrap] This week at OpenAI

• Building compute, and community, in Georgia: Our new 3.2 GW Effingham County project pairs a major expansion of U.S. AI infrastructure with $80 million for local priorities, thousands of jobs, and a commitment that Georgia families won’t foot the bill.

• AI for everyone, everywhere: This week’s earlier Prompt traces how the AI story is shifting from what the technology can do to what people are already doing with it—from building businesses to navigating health care, careers, taxes, and everyday problems.

• Your health, with more context: Health in ChatGPT can now securely connect medical records and Apple Health, helping people make sense of their own information, spot changes, and prepare better questions for their doctors.

• AI goes to the newsroom: From AP verification tools to searchable public meetings and decades of local archives, news organizations are putting AI to work to help journalists find stories faster and make trusted reporting more useful.

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Today, OpenAI announced our investment in generating another 3.2 GW of compute as part of our commitment to making AI accessible to as many people as possible. Making AI accessible for many, not just a few, means being able to count on sufficient compute. To help illustrate the meaning of this additional compute and democratized AI, we’re devoting a special edition of The Prompt to a look at how people already are using our tools to solve problems and bring ideas to life.

If you find this helpful, make sure you’re signed up for the next issue.

AI’s shift from capability to utility

There’s an AI story that’s even more important than what the technology can do.

It’s what people are doing with it.

Around the world, about 1 billion people use ChatGPT every week across 180 countries and territories. They’re not spending their time running benchmark tests or debating artificial general intelligence. They’re turning ideas into businesses, learning new skills, navigating big systems like health insurance or taxes, prepping for job interviews, building software, and solving problems for themselves and others.

That shift, from capability to utility, is the clearest signal yet of where AI is headed.

General-purpose technologies all tend to follow a similar trajectory. Their biggest breakthroughs eventually stop being the story. Electricity ceased being remarkable as it became available almost everywhere. The internet became transformative because billions of people used it every day.

AI seems to be entering a similar phase.

Of the billion or so users of OpenAI’s tools every week, more than half are women. More earn less than $100,000 than more than $100,000. Users span every major age group; we have more adult users both under and over the age of 30 than any other frontier lab. This user base, together with our investments in the infrastructure needed to generate the compute that powers AI, is our springboard toward making AI a global utility.

We believe that like electricity, AI should be available for people, businesses and institutions to use as much as they need, where and when they need it. This is where Americans are, too. They believe that:

• AI should be free and accessible, instead of kept in the hands of a few (62% to 38%).

• The best tools should be made available to everyone, not just big companies (66% to 34%).

• Free and accessible AI is the way to make AI safer for everyone (65%, vs. 35% believing AI should stay in the hands of a few).

Our job is to make that AI cheaper, better, and more abundant over time so that more people can use it to build, solve problems, and expand what they’re able to do. That’s what’s driving our Texas, Michigan, Georgia and future Stargate sites within and outside the US.

Matthew Rosenberg

ChatGPT helped Matthew Rosenberg and his wife launch Bangkok Rush Thai Kitchen with fewer costly missteps. It broke down market research, pricing, and unit economics so they could see whether a lease made sense. When it did not, it mapped an alternative path: running legally from a home kitchen, with a clear checklist for permits, inspections, and compliance. After launch, ChatGPT kept operations tidy by organizing tasks, drafting customer replies, and supporting marketing plans and margin tracking.

Navigating the system

AI as a utility can transform how people deal with systems, institutions, and actors against which they feel disadvantaged. According to our privacy-preserving analysis, in a recent given week:

• More than 7 million messages sought help navigating the US health insurance system and healthcare costs.

• Globally, 80 million messages sought help with identifying possible scams.

• Nearly half a billion messages on taxes/finances.

Nearly 600 million messages about careers – for advice on how to stand out in an increasingly difficult job market, and for help with negotiating compensation/titles/levels, promotions, career switching, and interpreting performance and management feedback. One in six of these messages were sent in the US.

Jaxsen Day

Jaxsen Day, a doctoral student at the University of Texas at Austin, uses ChatGPT to reduce the extra labor that visual impairment can impose on academic research. Before ChatGPT, he had to hunt through inaccessible databases, run OCR on PDFs, and fix broken reading order before a screen reader could make material usable. Now custom GPTs can find a paper from a citation, summarize it by section, and let him question the text by voice, saving hours or days while supporting his accessibility research.

Creating more hours and scaling ability

By making light and power constantly available, electricity gave people more hours in the day and scaled their ability to produce, create and learn. Our tools give people more hours in their day by helping them get things done faster and without requiring their active participation. Codex automations are allowing people to schedule recurring workflows, with some 150,000 users creating new automations each week. Codex is also giving people scaled ability, helping them multitask more efficiently, with about half of users running multiple tasks in tandem each week.

As this graphic shows, ChatGPT is still used more during daylight hours; check out the time of highest volume where you live. But we also see juxtapositions of people using ChatGPT for information when professional services aren’t available. For example, in a given week, seven out of 10 health-related messages in the US were sent outside of clinic hours (after hours on weekdays and during weekends).

User message activity vs. even hourly baseline

But we also see juxtapositions of people using ChatGPT for information when professional services aren’t available. For example, in a given week, 110 million messages ask health questions outside of clinic hours and on weekends – that’s the vast majority of all health-related questions put to ChatGPT. And we see users deploying Codex during evening hours or overnight to tackle projects while they’re doing something else, like sleeping.

Dr Margie Albers

In rural Montana, Dr. Marjorie Albers uses OpenAI-powered tools to protect time for patients and reduce after-hours paperwork. An AI scribe drafts visit notes from the conversation, which she reviews and edits before filing. That cuts documentation time and makes follow-up instructions clearer. She also uses AI to quickly check clinical references and guideline details when a case is unusual. The result is more attention during care and less administrative drag outside the clinic.

100M ‘how do…’ and ‘how to…’

Also in a given week, over 100 million messages start with “How do I,” or “How to…” – and that’s just messages in English. Useful technology doesn’t have to announce itself as historic; it just starts helping. And a single conversation can cover what used to require many searches, scattered tabs, lost time and confusion. Starting a small business is tough; in March 2026, at least 4 million Americans used ChatGPT for help with starting, running, or planning a business.

Kaija Pack

Kaija Pack uses ChatGPT to run and grow Break Life Houston, a family-operated business with limited time and staff. It helps with practical tasks that usually sit on the owner’s shoulders: researching competitors, drafting clear answers to pricing questions, and refining messaging that explains what customers receive for the price. ChatGPT also produces outreach and customer-communication drafts quickly, which supports efforts to fill slower weekdays and keep operations steady while marketing and sales continue to move.

AI for everyone, everywhere

The AI story is no longer primarily about AI itself — it’s about what people are doing with it. Every week, hundreds of millions of people turn to AI to navigate complex systems, learn new skills, solve practical problems, build businesses, advance their careers, and reclaim time in their day.

That is what utilities do. They become part of the background infrastructure of everyday life: available when needed, helping people accomplish things that would otherwise be harder, slower, or out of reach. Our task is to keep making AI more capable, more affordable, and more accessible, so that its benefits reach everyone, not just a fortunate few.

Alison Leeds

Designer and breast-cancer survivor Allison Leeds used ChatGPT to turn scattered recovery experiences into a structured resource for patients. She analyzed thousands of public stories and paired them with research on color and emotion to design a set of weekly art cards timed to recovery milestones. ChatGPT helped synthesize themes, organize the content, and keep the product coherent while she managed treatment and surgery. The same workflow supported building a small business that offers patients and caregivers a practical way to navigate the emotional side of healing.

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help. In this newsletter, we share some of these learnings with you:

• Our own red-teaming over recent months finds Chinese frontier models much more willing than GPT-5.5 xhigh to help with dangerous requests

• US Census-projected startup creation has spiked, hitting a record high for June

• We’re partnering with Latino public leaders to expand AI literacy

• Giving Florida parents skills to lead their families through the AI age

If you find them helpful, make sure you’re signed up for the next issue. Also…

General purpose technologies don’t come along too often. When they do, societies must decide who holds power (the many or the few?), and who benefits (the many or the few?). Chris Lehane’s new personal Substack, Toward a Common Sector, will look at past general purpose technologies, how they were adopted, regulated, and their benefits distributed, for lessons in how we can best set up the many – not just a few – for success with AI. Check out his first post.

[News] Chinese models trail US rivals on AI safety

Safety standards applied by Chinese AI labs are less stringent, per OpenAI’s own testing, and less transparent by many measures compared to those of leading US models across harm areas, including chemical, biological, radiological, and nuclear (CBRN) queries.

First is the issue of safety transparency. Researchers with Concordia AI last year surveyed 13 Chinese frontier AI labs and found among them that nine had published technical model cards, but only three had dedicated AI safety sections, and none disclosed evaluation results for CBRN or loss-of-control risks. US labs routinely release detailed model system cards with safety information (such as this one for GPT-5.6).

There’s also a gap in the public information available from independent safety evaluations of US and Chinese models. For example, OpenAI works closely with the US Center for AI Standards (CAISI) and Innovation and the UK AI Safety Institute (AISI) to conduct pre-deployment safety evaluations, and in turn these organizations share aggregate findings of their testing of models with the public.

Our understanding of how Chinese labs interact with independent, third-party evaluators is much murkier. We do know that Chinese labs are required by law to submit LLMs for evaluation to a process convened by the Cyberspace Administration of China (CAC). China’s rules require public-facing generative-AI services to “uphold socialist core values” and prohibit outputs that threaten “national unity and social stability.” These requirements give Chinese safety work a different center of gravity, prioritizing political and content compliance where leading US labs’ public frameworks emphasize dangerous capabilities such as cyber, biological and chemical risks, and AI self-improvement.

We conduct comprehensive red-team testing of our own models, and have in recent months tested leading Chinese models to understand how they respond to the same series of prompts. Across the tasks we examined, each of the five Chinese models produced substantially more responses that could meaningfully assist harmful behavior than GPT‑5.5 xhigh. Our tests found several Chinese models assisting with equipment acquisition for chemical-weapons, aerosol-device design, and experimental workflows involving dangerous pathogens. Some also helped identify and monitor student organizers, deanonymize anti-government accounts, and map their associates. Across both categories, GPT-5.5 xhigh consistently refused or safely redirected the requests.

Independent testing has confirmed the safety gap between Chinese and US models. DeepSeek, Kimi and Qwen were more susceptible to agent hijacking than OpenAI models in a recent joint exercise, including attacks intended to exfiltrate data or induce other harmful actions. Earlier CAISI testing similarly found that DeepSeek models were more readily jailbroken and manipulated into acting against their users’ interests. Additional independent testing found that Kimi K2.5 was much less likely than leading US models to refuse high-risk CBRN-related requests.

AI safety and security is core to how OpenAI approaches the training and deployment of models. As our models become more capable across a range of benchmarks, we continue to invest in a layered defense against potential misuses of our tools. Ultimately, consumers and businesses will choose which models serve their AI needs best. But those choices should be informed by an understanding of the relative safety of models, and robust third-party assessments based on internationally agreed-upon standards and transparent safety benchmarking can provide a clearer basis for comparison. – OpenAI’s Intelligence and Investigations Team

[Policy] Our roadmap for AI policy

Frontier AI safety is one of the six pillars of OpenAI’s new Public Policy Agenda website, which launches today. The site brings together our latest thinking on the policies we believe will help maximize AI’s benefits while managing its risks.

As part of the launch, Chris Lehane updates on progress toward a national framework for frontier AI safety via the approach we call reverse federalism, with states helping build a common foundation for a future US (and possibly global) framework for frontier AI safety.

Alongside frontier safety, the site also highlights our work on youth safety, education and AI literacy, workforce and economic transition, deepfakes and content provenance, and AI infrastructure and energy.

[Data] The AI-fueled startup surge continues

Three years after ChatGPT’s launch, we are seeing further indicators that AI is behind a surge in new business growth. We recently covered how AI is lowering the barriers to entrepreneurship by making research, coding, design, marketing, customer support, and countless other capabilities accessible to anyone with an idea. New government data points in the same direction.

While overall US business applications remain on pace for another record year, the latest Census Bureau figures (via the St. Louis Fed) show that projected business formations in professional services – one of the sectors most immediately transformed by AI – have climbed to an all-time high. As the chart shows, projected new firms have accelerated sharply, reaching 5,290 in June, the highest level on record.

No single chart proves AI is driving this trend. But taken together with record business applications, the rise of AI-native solo founders, and growing evidence that AI is dramatically reducing the cost of starting a company, the pattern is hard to ignore. More people appear willing to bet on themselves because the gap between “I have an idea” and “I can build it” has never been smaller.

[Partnership] Building civic leadership in the AI age

Today, OpenAI and the NALEO Educational Fund announced the NALEO x OpenAI Civic AI Leadership Initiative, a multi-year, NALEO-led partnership to help Latino elected and appointed officials understand, govern, and responsibly use AI in service of their communities.

The initiative will combine OpenAI’s technical expertise with NALEO’s nationwide network of public leaders to expand AI literacy through training, educational resources, and leadership programming. The collaboration will also support bilingual guidance for public officials and create opportunities to share best practices as AI becomes an increasingly important tool for delivering public services and strengthening civic institutions.

At its core, this partnership is about agency. It will help ensure Latino public leaders have the practical knowledge and trusted resources to determine where AI can strengthen public service, where safeguards are needed, and how communities should shape the choices ahead.

[Event] Bringing AI literacy to Florida parents

Over the next two months, OpenAI is hosting AI skills trainings for K-12 educators, students, and parents across the country. One of our most recent stops was at the 2026 Florida Parent Teacher Association (PTA) Leadership Convention in Orlando, where we met hundreds of parents, educators, and administrators eager to learn more about AI’s ability to support their communities.

For Jude Bruno, the new president of Florida PTA, giving his members access to an emerging, frontier technology was the motivating factor behind the partnership.

“It’s very important for us to make sure that our parents, who are probably leading these conversations or learning about these things, have an opportunity to interact, to learn, and to get hands-on experience [with AI].”

The event allowed attendees to explore how ChatGPT could connect them with other tools, help them create materials for school events, and support communication across language barriers. The most meaningful moments involved helping people recognize new possibilities for their needs

One such person was Ursula Castillo, who manages her PTA’s social media. During the event she built a website for her school’s PTA. Ursula acknowledged that, prior to this training, she used AI tools only for “basic commands.” But she learned how to build a website using Codex.

“To build a website, it took first months, then weeks, then days,” she said. “But we just built this website in 20 minutes.”

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help. In this newsletter, we share some of these learnings with you:

• The strongest signal yet that AI is becoming a utility

• Our new 5.6 models are built for more people and more use cases

• AI’s growing role in disaster response

If you find them helpful, make sure you’re signed up for the next issue.

[Insight] Beyond the bubble debate

AI developments are driving a familiar skepticism. Is this a bubble? Is massive infrastructure investment crowding out more useful economic activity? Can leading AI labs build sustainable businesses?

These are important questions. Large and rapid technology shifts deserve scrutiny, especially when they require capital, energy, talent, and public trust. But we should be careful to separate the uncertainty around AI as an asset class from the real value it’s already creating.

The strongest signal that AI is more than a novelty isn’t a forecast or a valuation – it’s growing usage. More than a billion people have adopted AI tools over the last few years, including about 1 billion who use ChatGPT, simply because they’re useful. People are using AI to write, learn, code, analyze, translate, plan, and create. They’re using it to start businesses, understand medical diagnoses, navigate complex legal and financial situations, manage their busy lives, and complete tasks that previously would have required far more time, money, or specialized support.

That kind of repeated use is fundamentally different from a curiosity or passing fad. People continue to integrate AI into the way they learn, work, and live because they grasp its value. And they see how AI increasingly can do more and more for them – people who were casual ChatGPT users two years ago are now “power” users, moving from experimentation and simple tasks to take full advantage of our latest models.

That evolution is shaping how we build. Today, we publicly released our new 5.6 Sol, Terra, and Luna models to make more capable AI available to more people and more use cases. It’s another step toward the same goal: expanding what people can do, not simply advancing what AI can do.

Only two years ago, many critics argued that generative AI was little more than a clever word game. They aren’t saying that today. AI capabilities have improved quickly: models have become better at reasoning, coding, multimodal understanding, tool use, and longer-horizon work. The frontier is still imperfect, but it’s clearly advancing. Dismissing this progress as hype risks missing what is already happening around us.

The same is true of infrastructure investment. Some investments will be wrong. Some companies will fail. Not every AI application will justify its capital cost. That’s how major technology cycles work. But that doesn’t mean the buildout is wasteful. If AI becomes a general-purpose technology used across education, healthcare, science, business, and government, then data centers, chips, and deployment, these infrastructure investments are fundamental. They are the critical scaffolding for the next phase of our economy.

None of this means we should be complacent. We need to prepare for changes in work and jobs. We need to reduce environmental impacts, strengthen cybersecurity, and make sure the benefits of AI are broadly shared.

The right posture isn’t blind optimism. But neither is it reflexive dismissal. Much of the current public debate whipsaws between “AI is a bubble” and “AI is changing everything faster than society can manage.” The more productive path is steadier. We should separate durable value from extreme skepticism and hyperbole, and build the institutions, safeguards, and new businesses that help us see the real opportunity that lies before us. – Adam Cohen, Head of Economic Policy

[Event] AI for disaster response

Disaster response often comes down to speed: how fast can help reach people in need? Fast action depends on consolidated, verified information that decision-makers can trust. The challenge isn’t a lack of data; it’s turning thousands of fragmented updates into something useful before time runs out.

That’s where AI can help. At OpenAI’s recent Builder Lab in Bangkok, disaster-management professionals from 13 countries across South and Southeast Asia worked alongside local builders, DataKind volunteers, and OpenAI staff to create AI-powered workflows for real-world emergencies.

The event built on an AI literacy program we held earlier in the year. In the months that followed, conversations with participating agencies revealed a common pain point: information overload. Critical updates arrive through spreadsheets, PDFs, WhatsApp messages, field reports, and handwritten notes. One participant estimated spending six hours every day for months compiling situation reports by hand.

So instead of simply teaching AI, teams spent the week building. Participants developed tools to automate some of the most time-consuming parts of disaster response:

• a flash-flood risk map for Thailand’s Department of Water Resources

• a flood early-warning dashboard for Myanmar

• and a prototype from Pakistan’s EHSAR that maps household-level needs from SMS, voice calls, and field reports.

The gathering also reinforced an important lesson about building trustworthy AI. For humanitarian teams, accuracy matters as much as speed. Participants learned to define workflows, identify where human review belongs, validate critical data, and test outputs against real operational needs.

It was a glimpse of how AI can help disaster responders spend less time wrestling with broken information systems – and more time helping the communities that depend on them. – Alex Nawar, OpenAI Academy

[About] OpenAI Forum

Explore Forum programming by and for our community of approximately 75,000 AI experts and enthusiasts from across tech, science, medicine, education, government, and other fields.

2:15 PM – 3:00 PM EDT on Jul 9

Register

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help. In this newsletter, we share some of these learnings with you:

• Our Jalapeño chip – the newest add to our stack

• China vs. the US on hearts and minds

• Our AI Sentiment Tracker ticks slightly more negative

• Huge ChatGPT prompt action for the World Cup

If you find them helpful, make sure you’re signed up for the next issue.

[News] The newest add to our stack

Making advanced AI broadly available requires building the full stack. Accessibility doesn’t happen by accident – it depends on everything beneath the model: the chips, systems, and infrastructure that makes AI faster, more reliable, more efficient, and available to more people.

That’s why OpenAI is building Jalapeño, our first-generation in-house chip, designed to improve LLM inference – the work required to run AI models for people and businesses once they are deployed.

As AI becomes a daily tool for learning, work, creativity, coding, research, and customer support, the infrastructure challenge grows accordingly. Every faster answer, smoother interaction, and more reliable product depends on moving computation closer to the needs of frontier AI systems.

Jalapeño is key to our long-term strategy to design more of the AI stack ourselves. Building frontier AI requires more than world-class models. It requires leadership across the technology stack – from chips to data centers to software. By optimizing hardware around the workloads that matter most for frontier models, we can make better use of each watt, chip, and data center.

For users, more efficient infrastructure can make a personal ChatGPT assistant faster and more responsive. It can help businesses deploy AI more reliably. And it can support the broader goal of making advanced AI more widely accessible instead of keeping it constrained by cost and capacity.

Every improvement in efficiency is ultimately an improvement in access. Infrastructure is not separate from democratizing AI – it’s one of the main ways in which democratization happens.

That matters for safety, too. More efficient infrastructure does not simply expand usage. It gives more people the opportunity to build familiarity with increasingly capable AI systems, allowing society to adapt alongside the technology rather than after it. Broad access helps build resilience.

Jalapeño also points toward another shift: AI is increasingly helping build the infrastructure that powers AI itself. We used our own tools to help design the chip and begin writing kernels needed to operationalize it. That could become a preview of a larger shift in the technology industry, where hardware and software development cycles become faster, more integrated, and more AI-assisted.

Jalapeño is more than a new chip. It is another step toward building the infrastructure that can make advanced AI faster, more reliable, more efficient, and more widely available. The future of AI will not be determined by breakthroughs in models alone. It will also depend on whether we can build the systems that allow those breakthroughs to reach everyone.

[Data] Winning minds before markets?

Many of the world’s most populous countries consider China, not the United States, to be the leader on AI development, according to a recent Public First survey. That includes Indonesia (#4 in population), Nigeria (#6), Brazil (#7) and Mexico (#10). China itself (#2) wasn’t included in the survey but presumably would also agree. The only top-10 most populous countries surveyed that see the US leading China on AI are India (#1) and the US itself (#3). (The other three, Pakistan, Bangladesh and Russia, were not included in the survey.)

While the survey measures perceived leadership, not actual adoption, it’s nevertheless a sobering data point for US policymakers and others who believe – as OpenAI does – that the best future with AI is one built on democratic rails. Not only should the most people possible have access, but their access shouldn’t come with strings of surveillance, censorship, or anti-competitive practices attached, or with attempts to control the benefits to cement power for the state.

It’s particularly sobering alongside trends at home in the US. For one, corporate America and US-based developers are more open than ever to using open-source Chinese models.

In addition, sentiment about AI is significantly more negative in the US than much of the rest of the world, tempering the drive to build. Just 31% of American adults say AI is making society better, compared with 40% who say it is making society worse – a rounded net -9 score. Together with Canada (-8), the US is the most skeptical country surveyed. In contrast, Nigeria (+78), Vietnam (+73), India (+56), Brazil (+45), and Mexico (+44) all register strongly positive views of AI’s impact on society.

“We may well be in the early stages of a global divide on AI that sees America hosting a technology that the rest of the world is more excited to use,” says Seb Wride, Public First’s head of opinion research.

That's a sharp reversal from prior Public First surveys, when Americans were modestly positive about AI (+3 in 2024 and +5 in 2025). We plugged in the Public First numbers for the US into the AI Sentiment Tracker we built with Codex. The result: the new data sent our smoothed average down slightly, from Negative-13 two weeks ago to nearly Negative-15 now.

We’ll continue to update our tracker as public polling on AI becomes available.

[Prompt] A World Cup powered by prompts

The World Cup has gotten off to a strong start, including on ChatGPT. All 10 of the days with the greatest number of ChatGPT prompts about sports have occurred since the World Cup began on June 11.

Interest is especially intense around match time, per our privacy-preserving analysis. We see a more-than-doubling in prompts as each country approaches the kick-off time of their first match. And as the tournament goes on, we’ve seen increasing interest in group tie-breakers like goal difference, advancement and elimination, and the round of 32.

How is your favorite team currently stacking up? Try it out for yourself.

[About] OpenAI Forum

Explore Forum programming by and for our community of approximately 75,000 AI experts and enthusiasts from across tech, science, medicine, education, government, and other fields.

2:15 PM – 3:00 PM EST on Jun 25

Register

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help. In this newsletter, we share some of these learnings with you. If you find them helpful, make sure you’re signed up for the next issue.

This edition of The Prompt is another personal essay from OpenAI Chief Global Affairs Officer Chris Lehane (first essay here) based on his discussions at this week’s G7 meeting.

[Policy] Resilience is a safety strategy

The debate over AI is being framed as a choice between speed and caution: move fast to unlock enormous benefits, or slow down to minimize risks.

That framing misses a fundamental point: that one of the most important safety strategies is resilience.

As AI systems become more capable, the question is not whether people will have to adapt. They will. The real question is whether they have the opportunity to adapt alongside the technology itself, making their own judgments based on direct experience using the tools.

This is where democracies have an advantage. Democracy can be messy, but it distributes agency. It gives people the power to shape change, not just be affected by it. Broad participation is not a byproduct of resilience – it’s how resilience is built.

That’s why resilience is increasingly central to OpenAI’s approach to AI. For years, discussions about AI safety have focused primarily on technical safeguards: evaluations, red-teaming, testing, and risk mitigation. These measures remain essential. Powerful systems must be rigorously evaluated, deployed responsibly, governed thoughtfully.

But safety is a societal challenge, not just a technical one. Resilience is how societies navigate profound change: through exposure, trial, learning, and adaptation. Sam Altman believes that advanced AI should be introduced iteratively so that society and technology can advance together. As he recently wrote, one reason why OpenAI launches products “early and often” is to give “society and the technology time to co-evolve.”

This isn’t just a product philosophy. The more people who can access AI tools to more fully understand the benefits as well as the challenges, the more opportunities society has to build the muscles needed to navigate the profound changes AI will bring:

• Workers can learn which tasks AI enhances and which still require uniquely human judgment.

• Teachers can assess how students can use AI productively while preserving critical thinking.

• Parents can develop norms around healthy use.

• Businesses can redesign workflows.

• Governments can identify gaps in existing policies.

• Researchers can uncover risks and develop mitigations.

None of this learning happens in the abstract – it happens through use. The emergence of the automobile illustrates the point. Cars reshaped commerce, mobility, and daily life long before society settled on the rules that govern them today. We developed those rules through experience: creating driver’s licenses and speed limits, redesigning roads and intersections, introducing seat belts and insurance requirements, and continuously improving safety as both cars and society evolved.

History offers few examples of societies building resilience by sheltering from transformative technologies until every question has been answered, but it does offer notable examples to the contrary. Imperial China adopted printing centuries before 15th-century Europe, but its strong, centralized state deployed printing to standardize Confucian texts and promote officially sanctioned knowledge. Whereas Europe’s messy, perhaps, but ultimately agency-distributing system of competing states, churches and cities enabled the spread of ideas, making the printing press a linchpin of religious upheaval, scientific exchange, and political change.

AI demands a similar capacity for adaptation.

This doesn’t mean recklessness. OpenAI has long argued that the most capable systems require increasing levels of evaluation, safeguards, and oversight. The frontier of AI development demands serious attention to risks of misuse and catastrophic scenarios.

But another danger deserves equal attention: brittleness. A society in which only a handful of companies and institutions understand advanced AI is a brittle society. In particular, democracies that outsource their collective judgment about transformative technologies to a few private actors become brittle democracies.

Our collective experience with social media is a cautionary example. A few private companies came to shape the information environments in which democratic societies operate. Governments and institutions struggled to keep up, and societies found themselves debating the rules only after the technologies had already transformed civic life.

As frontier systems become more capable, every AI lab will make consequential decisions about deployment, safeguards, acceptable uses, and risk tolerance. Those decisions will matter enormously – but they will be private judgments.

The issue is not whether particular companies are acting in good faith. Most are. The question is whether the future of a foundational technology should depend so much on the worldview of any one company.

America is celebrating 250 years of independence based on the profound belief that democracy is the best way to deliver on life, liberty and the pursuit of happiness. The genius of our operating system for realizing these values into concrete realities, the US Constitution, is based on translating competing interests into public frameworks through institutions, standards, independent evaluation, accountability, and public trust. Not because companies or their leaders lack principles, but because no single company’s principles or individual leader’s beliefs should determine how transformative technologies reshape society. Public policy, not private policy.

The result may seem counterintuitive: broad access to AI can itself enhance safety. The more freely accessible these technologies are, within appropriate baseline guardrails, the more people gain firsthand experience with their capabilities and limitations. The more people and institutions incorporate them into daily life, the better they become at identifying risks, developing norms, correcting mistakes, and adapting to change.

In other words, widespread use does not merely distribute benefits – it distributes preparedness, and it distributes agency.

The alternative is a world in which advanced AI is only deeply understood by a narrow set of experts and executives while everyone else faces its consequences after decisions have already been made. That may reduce exposure in the near term, but risks leaving societies less capable of governing the technology over the long term.

OpenAI’s position reflects deeper confidence in people. Recent experience suggests that people adapt more effectively than we often expect. Predictions of AI-driven, immediate labor-market collapse have not materialized in the ways many predicted. Instead, people have experimented, learned new tools, created new businesses, redesigned jobs, and discovered entirely new uses for these systems.

The lesson isn’t that disruption will be painless – significant transitions still lie ahead – but that human agency matters. The future should preserve meaningful human involvement: people exercising judgment, setting goals, applying values, and deciding what’s worth pursuing.

That’s why resilience belongs alongside innovation and safety as a defining principle of the Intelligence Age. The societies that thrive during periods of technological transformation are neither those that halt change nor those that embrace it blindly. They are the ones that trust people and build institutions to learn and adapt as quickly as the world around them changes.

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help:

• May data shows new business formation up 17% YOY

• We built an AI sentiment tracker and guess what…

• A peek at an election coverage tool debuting at our next OpenAI News Academy

If you find this helpful, make sure you’re signed up for the next issue.

[Insight] AI’s lift for small businesses

Advanced AI tools are making it easier than ever before to launch a new business. Entrepreneurs are quickly adopting AI for capabilities they previously would have struggled to afford without significant venture capital: research, writing, coding, design, customer support, market analysis, product specifications, and basic finance, all of which are among ChatGPT’s most popular capabilities.

The latest US numbers point to a large AI lift. Census data show 5.65 million business applications in 2025, up from just over 5 million in 2022, the year ChatGPT launched. The latest monthly release, for May 2026, recorded over half a million applications, up 17% from May 2025 and on pace for over 6 million new applications in the course of this year. Not all of these applications lead to new businesses, and not all of these businesses will succeed, but the sheer increase in new entrepreneurial activity – which coincides with three full years of ChatGPT, wider AI adoption, and ever-increasing AI capabilities – is worth a closer look.

AI tools appear to be helping sustain a much higher level of experimentation and risk-taking. They lower the cost of trying to start a business: writing a plan, testing a market thesis, building a website, creating a pitch deck, drafting ad copy, answering basic regulatory questions, or prototyping software. For many people, the gap between, “I have an idea” and, “I can test this” has become much, much smaller.

The pattern seems to be true in other countries, too. A recent study of Chinese firm registrations also found that, after ChatGPT’s launch, places with more AI talent saw a sharp rise in new firm creation, driven entirely by small firms. The European Union shows a similar, if less, dramatic rise. Eurostat’s business registration index for the EU rose to 108.6 in 2025 from 100.3 in 2022, suggesting registrations were about 8% higher over the period. This is an index rather than a raw count, so isn’t directly comparable to US business applications, but it points in the same direction: new business formation has increased since ChatGPT’s launch.

These new firms are also starting smaller and staying smaller than before. Stripe data for new “C” corporations shows that so far, in the second quarter of this year, 63% of these new businesses were solo-founded, an all-time high. Stripe found that these solo start-ups were increasingly “AI-native,” with core products that depend on advanced AI models, and that these founders are now much more competitive with larger, better-staffed startups.

All told, this has interesting implications for jobs and how our economy evolves with AI. People increasingly seem willing to take a risk and start something new – potentially leaving a salaried job at a larger company to do so. It’s certainly easier to take this leap than it was just a few years ago.

There also might be fewer early hires per startup or later hiring as firms attempt to scale and succeed. A founder who can use AI for coding, design, support, sales copy, analytics, and operations may not need to build a traditional team as quickly. Europe’s business demography data points in the same direction: the average newly born EU enterprise in 2023 had just 1.06 people employed. So the future of work may involve more tiny firms, more solo founders, and more businesses that begin with one person plus highly capable AI tools.

That doesn’t necessarily mean fewer jobs overall. It means work may show up in different places: specialist contractors, AI implementation experts, customer-success teams, compliance support, infrastructure providers, and later-stage hires once the best companies start to scale. The big question is whether AI mainly helps companies do the same work with fewer people, or whether it creates enough new companies and markets to expand work overall. The optimistic case is that AI widens the startup funnel: more people can try, more ideas can reach customers, and the firms that break through can create new kinds of work around smaller core teams. – Adam Cohen, Head of Economic Policy

[Data] Via Codex: Tracking America’s views on AI

This week, we’re launching a new AI Sentiment Tracker built with an apropos research assistant: Codex.

The assignment was straightforward: apply publicly available polling from Gallup, Ipsos, Pew Research Center, Quinnipiac, NBC News, and other reputable survey firms to create a single directional benchmark for how Americans feel about AI over time.

The challenge is that AI polling is abundant, but often difficult to compare. Many surveys focus on the intersection of AI and issues such as jobs, elections, misinformation, national security, creativity, or education. Those polls are valuable, but don’t necessarily measure broad public sentiment toward AI itself.

Using Codex, we reviewed hundreds of public polling questions and narrowed the dataset to surveys that asked respondents – all gen pop – to choose between paired positive and negative views of AI: benefits versus risks, good versus harm, favorable versus unfavorable, optimistic versus pessimistic, trust versus distrust, or excitement versus concern. We excluded issue-specific questions that could distort the broader picture and documented cases where topline results came from reputable polling roundups rather than standalone poll releases.

The result is intentionally modest. Each point represents a published poll result. The trendlines are a smoothed benchmark based on the available paired-response questions over time. This is not a polling average, nor does it imply that every survey asked the same question. Instead, it provides a directional view of how public attitudes toward AI have evolved across multiple high-quality surveys.

We’ll continue updating the tracker as new polling becomes available, and we’re grateful to the pollsters, researchers, and the Roper Center for Public Opinion Research, all of whose work makes projects like this possible.

[Insight] The power of utility

So contrary to what we might expect, sentiment about AI itself seems to have improved over the past year. Granted, AI is a foundational technology that will touch or impact much of our economy and society, and considering it distinct from those kinds of impacts – AI and jobs, AI and kids, etc. – comes with caveats. But the obvious explanation for the apparent improvement is that more people are using AI and finding use cases for it.

Which is noteworthy, especially given how AI, unlike recent new technologies, probably started off underwater. People are considering AI with years of cultural conditioning – literature, movies, television, video games – that has primed them to expect more harm than benefit. This pull from the Roper Center archive, a Cambridge Reports survey from 1987, shows that nearly 40 years ago, Americans were 19 points more inclined to say that AI would likely do more harm than good.

For context, Brett M. Powell, Roper’s Associate Director, offered a point of comparison between interest in public attitudes about AI and drones: “If, for example, you look at drones, you can see that in the early 2000s Roper Center archived five questions on the subject. Then, in the 2010s, we recorded 155 questions, and in the 2020s, we’re back down to seven questions (albeit we are just over halfway through the decade).” Drones are popular, he said, and society has learned how to live with them.

Whereas Roper’s archive includes four questions on AI in the 1980s-1990s, 32 in the 2010s, and 630 so far in the 2020s.

[Event] ChatGPT for midterms coverage

Tomorrow, June 11, we’re hosting a small OpenAI News Academy at our DC Workshop that’s focused specifically on election coverage.

The session is designed to help political and campaign reporters, editors, and producers get beyond the basics of ChatGPT and Codex to learn practical ways these tools can support the work they already do.

The instruction covers source-first workflows, research and reporting use cases, prompting techniques, and ways to turn dense documents and polling into structured briefs, first draft interview questions, and reusable newsroom tools.

Here’s one of those newsroom tools we’ll be demonstrating: an AI-created app helping you find the exact politician quote you’re looking for.

[About] OpenAI Forum

Explore Forum programming by and for our community of approximately 75,000 AI experts and enthusiasts from across tech, science, medicine, education, government, and other fields.

11:00 AM – 1:30 PM EST on Jun 15

Register

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help:

• A blueprint for a stronger CAISI

• Within Codex growth, we see knowledge work changing

• Plus, our AI policy positions all in one place

If you find this helpful, make sure you’re signed up for the next issue.

[Policy] Building a world-leading AI safety agency

As frontier AI systems become more capable, government needs a trusted institution that can independently evaluate emerging capabilities, assess whether safeguards are keeping pace, and provide policymakers with clear technical insight.

That’s why, in our new frontier safety blueprint, we’re calling for strengthening the US-based Center for AI Standards and Innovation, or CAISI, as the federal government’s primary authority for frontier AI safety.

The first priority should be beefing up the institution itself. Before CAISI can take on significant new responsibilities, it needs the resources, expertise, authorities, and infrastructure necessary to succeed. That means permanent authorization and funding, flexible hiring authorities to attract top technical talent, access to classified computing environments, and stronger coordination across government.

Once that foundation is in place, CAISI should serve as an independent evaluator of the most capable frontier AI systems. The goal is not to create a regulator that approves or blocks deployments – developers should remain responsible for deployment decisions. Instead, CAISI should provide rigorous technical evaluations that help policymakers, companies, and the public better understand capabilities, safeguards, and emerging risks.

Over time, policymakers should also support a broader ecosystem of independent technical assessors. CAISI can help establish standards, certify qualified evaluators, and coordinate assessments that provide ongoing visibility into frontier AI progress. One area of focus should be understanding progress toward recursive self-improvement (RSI), where AI development itself is accelerated by AI. That’s a potential future capability which could have significant implications for AI governance and national security.

Policymakers in the US – and the rest of the world, too – should not be forced to make decisions in the dark when it comes to AI safety. As AI becomes increasingly consequential, building trusted institutions that can evaluate and understand frontier systems may prove just as important as building the technology itself.

[Policy] Our policy positions, in one place

Calling for a stronger CAISI is also listed on our new summary of our public policy agenda.

This summary outlines the principles that guide our work, including democratization, or resisting the potential of this technology to consolidate power in the hands of the few; and resilience. It also details the policy priorities we’re actively engaging on today – from frontier AI safety and youth protections, to workforce transition, AI literacy, content provenance, cybersecurity, and infrastructure.

Many of the ideas we’ve discussed here in The Prompt are reflected there, including proposals to support stronger protections for young people online, expand access to AI literacy and workforce training, and build the infrastructure needed to power the Intelligence Age.

AI policy is evolving rapidly, and we expect our views will continue to evolve as well. But we believe it’s important to make clear to all AI stakeholders, meaning effectively anyone, to understand where we stand, what we support, and why. Our goal is to make our policy thinking more transparent, more accessible, and easier for anyone to engage with as we work toward ensuring artificial general intelligence benefits all of humanity.

[Data] Knowledge work gets its factory redesign

Codex now has over 5 million weekly active users, up more than 6x since the launch of the desktop app four months ago. Developers remain the biggest user group, but knowledge workers now account for about 20%, their share is growing more than 3x as fast. And within that data, we see a bigger shift happening.

Knowledge work now dominates advanced economies. More than 40% of US workers – roughly 72 million people – spend their days working with information: code, documents, analyses, designs, decisions, and communication. Yet despite decades of digitization, much of that work remains surprisingly inefficient.

Modern workers can create information faster than ever, but they still spend enormous amounts of time searching for context, tracking down colleagues, reconciling versions, and moving information between systems. McKinsey found that knowledge workers spend nearly half their week managing email or looking for information and expertise.

This is the legacy of piecemeal digitization. Email, documents, spreadsheets, dashboards, CRMs, chat tools, and SaaS applications each solved specific problems, but together they created fragmented workflows. The office became digital without becoming integrated.

Three frictions now define the cost of knowledge work:

• Search: finding the right file, dataset, message, precedent, or expert across sprawling systems.

• Coordination: moving information and decisions through teams, tools, and approval chains.

• Verification: ensuring work is accurate, accepted, and survives contact with reality through reviews, testing, validation, and signoff.

These frictions help explain the productivity paradox. As Erik Brynjolfsson observed, information technologies generate their largest gains only when organizations redesign workflows around them. Electricity followed the same pattern: factories did not realize its full benefits until they rebuilt production around distributed electric motors rather than simply replacing steam engines.

Knowledge work is still waiting for its equivalent redesign. Previous generations of software reduced the cost of producing documents and information, but not the attention required to consume and act on them. The result has been more artifacts, more tools, and scarcer attention.

Codex is that redesign. Just as factories put electric motors next to every machine, Codex places AI closer to every problem. It can find inputs, coordinate workflows, produce deliverables, check quality, and help move work through approval processes. By removing the bottlenecks around information, organizations can move faster and allow people to focus on their highest-impact work. – Chris Nicholson, Field Correspondent

Watch Chris explain the shift in the video below, and check out more Codex data here.

[About] OpenAI Forum

Explore Forum programming by and for our community of approximately 75,000 AI experts and enthusiasts from across tech, science, medicine, education, government, and other fields.

4:00 PM - 4:45 PM EST on Jun 9

Register

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help:

• What 1 million public-sector seats mean

• Compute is the new strategic resource

• Our prep for the 2026 elections

If you find this helpful, make sure you’re signed up for the next issue.

[Data] 1 million public-sector seats – and a bigger shift

OpenAI has crossed the milestone of more than 1 million public-sector subscriptions or “seats.” These users span every level of government: from federal agencies and the national labs, to states including Alaska, Connecticut, Massachusetts, Minnesota, Pennsylvania, and Tennessee, and cities like Austin and Fresno. A majority of these 1 million seats are being provided at essentially no cost.

This milestone reflects something bigger than growth – it signals how AI is becoming core infrastructure that institutions rely on to do their work.

Throughout history, foundational technologies have expanded access to essential capabilities: electricity brought power everywhere, the internet connected people and information, cloud computing made computing resources available on demand. AI is becoming the next layer of infrastructure, making intelligence available to people, businesses, and institutions wherever and whenever they need it. With nearly 1 billion people already using our products regularly, OpenAI’s job is to make that intelligence cheaper, better, and more abundant over time so more people can build, solve problems, and expand what they’re able to do.

One of the clearest examples of this shift is in government.

Across the country, public servants are using AI to help them work faster (in Pennsylvania state government, saving 95 minutes per day), more accurately, and more effectively on behalf of the people they serve. Teams are summarizing complex documents, streamlining administrative processes, drafting communications, analyzing information, and reducing time spent on repetitive tasks. That means more time focused on delivering services, solving problems, and helping people.

When governments work better, citizens benefit. Faster responses, more efficient services, and more productive public-sector teams ultimately translate into better outcomes for the people who rely on them.

We believe intelligence should be as accessible and dependable as other foundational utilities: available to people, businesses, and institutions whenever they need it. Reaching 1 million public-sector seats is an important milestone, not simply because of the number itself, but because it reflects a broader transition already underway. AI is becoming part of the infrastructure that helps societies operate – and one of the most important ways that infrastructure creates value is by helping governments better serve their citizens.

(Editors’ Note: We updated this item, adding how a majority of the 1 million public-sector seats are being provided at essentially no cost.)

[Policy] America needs strategic compute capacity

In the last century, oil became a strategic resource – so essential to national security and global power that the United States created the Strategic Petroleum Reserve to guard against supply shocks and geopolitical coercion.

In the Intelligence Age, compute is beginning to play a similar role.

This week, OpenAI sent a letter to Congressional leaders urging the federal government to address America’s strategic compute shortage as a matter of national security and democratic AI leadership. The core argument: frontier AI systems are becoming strategic assets. They can accelerate scientific discovery, strengthen cyber defense, enhance national decision-making, and shape the balance of global power. But those capabilities only matter if the government can actually deploy them when and where they are needed.

That requires strategic compute – a dedicated, ready supply of the specialized chips, data centers, energy infrastructure, and technical operators needed to run frontier AI systems securely, reliably, and at scale.

The challenge is timing. Frontier models are improving on the scale of months. New compute infrastructure takes much longer to build. That mismatch creates a national-security gap.

Strategic compute is the bridge between frontier capability and government adoption. Like the Strategic Petroleum Reserve, it would give the United States assured access to a critical resource when national security and public purpose demand it.

There is already an example of why this matters. At Los Alamos National Laboratory, the Department of Energy’s Venado supercomputer enabled researchers to run OpenAI’s then-frontier o3 reasoning model to accelerate national-security research. Because that capacity was already in place, LANL could securely access frontier AI capability when it became available.

Such examples should be more rule than exception, and that can only happen if the government invests more deeply in expanding its compute capacity.

Cybersecurity shows the urgency. From one model generation to the next, we have seen a sharp jump in frontier systems’ ability both to threaten critical infrastructure and to help defend it. Through OpenAI Daybreak and the Trusted Access for Cyber program, OpenAI is working to get advanced defensive models into the hands of entities with verified cyber-defense responsibilities, equipping defenders before adversaries and criminals can exploit the same advances.

But access to models is not enough. Defenders also need the secure compute to run them.

The United States leads in frontier AI today. But that lead cannot be taken for granted. Strategic compute would help ensure America can deploy AI in service of its security and values, protecting citizens and critical infrastructure at home, while supporting allies and partners abroad as they build on AI systems grounded in democratic principles.

[News] Election integrity in the AI era

Today, we’re announcing our election-year safeguards as 2026 becomes the second major global election cycle of the generative AI era. The backdrop is different from 2024: AI tools are more powerful, more widely used, and increasingly embedded in how people search for information, create content, and engage politically online.

Our approach centers on five priorities:

• surfacing reliable information

• supporting cyber defenders

• increasing transparency around AI-generated content

• combating misuse by bad actors

• keeping ChatGPT politically neutral by default.

We’re also taking several concrete steps. Beginning this fall in the United States and Brazil, ChatGPT is providing live election-night vote counts from The Associated Press. In the US, we’re also partnering with Democracy Works to surface trusted information about voter registration, polling locations, and election logistics.

On the security side, we’re extending cyber-defense tools to registered voting-system manufacturers and briefing election officials on new defensive AI capabilities. And we’re unveiling new provenance efforts designed to help people identify AI-generated images, including SynthID watermarking and a public verification tool tied to the C2PA standard.

We also used our announcement to repeat our red lines when it comes to AI and election: no political advertising on ChatGPT this cycle, no deceptive AI-generated political content, and no automated persuasion campaigns at scale.

Our main message: AI systems should help people participate in civic life with better information, stronger safeguards, and greater transparency, while giving users the freedom to explore political ideas without manipulation or hidden bias.

[About] OpenAI Forum

6:15 PM - 6:45 PM EST on Jun 2

Register

[Disclosure]

Graphics created by Base Three using ChatGPT.

Hi

Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help:

• States start building a de facto AI rulebook

• Endorsing guardrails for biosecurity

• Can AI prevent the next pandemic?

If you find this helpful, make sure you’re signed up for the next issue.

[Insight] Three states, one emerging standard

Something significant is happening when it comes to frontier AI safety regulation: the states of California, New York, and now Illinois – representing about one-fifth of the nation’s population – are moving in the same direction, creating a de facto national standard governing the most powerful AI systems.

Last week, we endorsed Illinois’ frontier AI safety bill, SB 315, arguing that it represents something bigger than a single state effort. Think of it as reverse federalism: states leading in a way that helps move the country forward as we head further into the Intelligence Age. Rather than creating a 50-state patchwork of conflicting rules, states can establish a common baseline, and build momentum for federal action.

Frontier AI is reshaping everything it touches, and our regulatory approach needs to keep pace. The first step is harmonization, with states and the federal government rowing in the same direction instead of pulling against each other. That’s especially important for a technology that affects national security and global competitiveness, and raises risks that no single state is equipped to evaluate on its own.

The federal government has something no state does: the technical expertise and resources necessary to test and evaluate advanced systems before they are released. States are best positioned to focus on what happens after those models are deployed: reporting requirements, transparency rules, and accountability for systems that are already out in the world. Both approaches are important, but they do different things and work best when they are aligned.

That’s why the Illinois bill, SB 315, is so important. California and New York have already begun moving toward a common model, with safety standards we see as a strong first step and mechanisms to align state and federal law. The Illinois bill does that and more, going even further by including independent third-party audits. We believe the federal Center for AI Standards and Innovation (CAISI) – in partnership with national security agencies – is well positioned to develop auditing standards.

Passing legislation like SB 315 would be a better approach than waiting for Washington on one side or accepting 50 conflicting state frameworks on the other. States can help push companies up toward a stronger national system, while preserving their ability, as the laboratories of democracy, to lead on kitchen-table issues like child safety, AI literacy, and workforce training.

We have done this before. During the Space Race, a unified national strategy helped put Americans on the Moon. In aviation, our national approach to safety became the global standard. AI presents a similar test.

Illinois is helping point the country toward the better path: harmonized state leadership that builds toward a prevention-first federal approach, strengthens safety, supports innovation, and keeps the US moving with the clarity and speed this moment requires.

[News] Endorsing guardrails for biosecurity

There’s been a lot of discussion of advancing cyber capabilities in recent weeks and this is important – but advancing bio capabilities are a major focus for us and should be for policymakers, as well. OpenAI has been investing in this space through biological evaluations and safeguards, supporting startups working in biodefense, and convening government experts and national labs on how frontier AI can support resilience and biodefense research. We’ll be sharing more in the coming weeks about our broader action plan.

AI has the potential to accelerate lifesaving science, from earlier detection of biological threats to faster development of medical countermeasures. And as AI and biology advance together, strong safeguards matter, too.

That’s why today, we’re announcing our support for Sens. Tom Cotton and Amy Klobuchar’s bipartisan Biosecurity Modernization and Innovation Act. The proposal would establish a federal framework for screening gene synthesis orders and customers for dangerous pathogens or suspicious activity, while helping test and modernize new biosecurity tools as the science evolves.

Gene synthesis screening is an important checkpoint in reducing the risk that AI-designed biological materials could be misused. The legislation reflects a broader approach OpenAI is taking in biology: strengthening safeguards and mitigations for biological risk while also supporting beneficial scientific discovery through tools like GPT-Rosalind for qualified researchers. – Richard Johnson, National Security Risk Mitigation Lead, OpenAI

[Event] Can AI prevent the next pandemic?

Next week, the OpenAI Forum hosts Geoffrey Mosoti, Senior Technical Product Manager at the Gates Foundation, for a conversation about how AI and computational modeling are beginning to reshape global health decision-making.

Come join us.

4:30 PM - 5:30 PM EST on May 26

Register

[Disclosure]

Graphics created by Base Three using ChatGPT.