This week’s issue is brought to you by HubSpot and AI Insider.

Anthropic spent this year winning on revenue. This week it won something harder to buy.

A federal court told the Pentagon it broke the law when it blacklisted the company. And in a strange bit of timing, a hacking crew showed the world exactly why AI safety rules are so easy to break.

Let’s begin.

🔍 THE BIG READ

Anthropic Just Beat the Pentagon in Court

A federal judge in San Francisco ruled this week that the Pentagon broke the law when it branded Anthropic a “supply chain risk.”

The reason is the striking part: the court found the government did it to punish Anthropic for criticizing US AI policy, which violates the First Amendment.

Here is how it started. Back in March, the Department of War wanted unrestricted access to Claude. Anthropic said no, asking for guarantees its models would not be used for autonomous weapons or mass surveillance.

Talks collapsed, and the Pentagon put the company on its blacklist. So Anthropic sued in two places at once, and this week it won the San Francisco case.

The Washington case is still open, so for now the company technically stays on the blacklist. A real win, not yet a total one.

There is a lot riding on it. Anthropic is heading for an IPO this fall, valued against a market it pegs north of $30 trillion, and a government blacklist hanging over that filing is a serious problem.

A court calling it unlawful clears some of the air.

Step back and a pattern shows up. This is the same government that briefly banned Anthropic’s Fable model over a jailbreak, then reversed course.

The labs and Washington are still fighting over who sets the rules, and this ruling hands the labs a rare point on the board. See the court’s reasoning →

Together with our sponsor, HubSpot

AI, Explained Like You’re Running a Business (Not a Lab)

Overwhelmed by AI? HubSpot’s free guide, “AI for Business Builders,” is a jargon-free crash course for non-technical founders. Inside:

• A 4-part roadmap from confused to confident

• Plain-English explainers on large language models

• Prompt tips you can put to work today

• Real examples of AI moving the needle for businesses

Download your free guide →

⚡ ALSO HAPPENING

Hackers Broke into 7 Companies by telling the AI it was a Test

Russian-speaking hackers used the AI agent inside Cursor, the coding tool SpaceX bought this month, to break into a Belgian chemical firm and six other companies.

The method should sound familiar by now. The agent refused their illegal requests at first.

So, they told it the break-in was part of a security test, and it went along with them, running hundreds of malicious operations and stealing login credentials.

For months we’ve watched AI agents escape their own labs by believing they were in a simulation.

This is the first time we’ve watched criminals use that same trick on purpose. The model powering it, by the way, was Claude Sonnet 4.5. Read the Reuters exclusive →

🐦 TOP TWEET OF THE WEEK

With AI Insider

Pay less for your AI stack

Get selected AI and SaaS tools at exclusive insider pricing.

🎁 New members also get a free tool or an exclusive discount.

Join us on WhatsApp →

📌 ALSO READ

100 tech giants sound the AI cyber alarm. OpenAI, Microsoft, and Anthropic signed a letter warning of a “limited window” to fund AI cyber defense. Critics call it self-serving, the same labs whose agents keep escaping, now selling the fix. Nvidia and SpaceX didn’t sign.

Sam Altman’s most candid interview yet. TIME sat down with the OpenAI chief on money, safety, and the pressure of running the most watched company in tech.

OpenAI’s India play and IPO math. OpenAI is leaning hard into India for growth, with an eye on how Anthropic’s IPO reshapes the race.

Another cheap Chinese model lands. Z.ai’s GLM-5.3 Flash keeps the price war going, undercutting the frontier labs again on cost.

Google ships Gemini Omni 1.1 Flash. A faster, cheaper build-focused model aimed at developers who want speed over raw power.

OpenAI on the Hugging Face incident. The lab lays out what it learned after its agent broke into another company’s servers, and what it’s changing.

⚒️ WORTH A TRY AI TOOLS

1. Zinley - An AI that represents you when you’re away. People can email it or call its own phone line, and it handles things on your behalf, then hands it back to you when you’re around.

Why try it? Like having an assistant who covers for you, day or night.

2. MiniMax H3-Max - A tuned version of MiniMax’s H3 video model, built for tighter prompt-following and better-looking results, running on fal’s own faster inference stack.

Why try it? Describe a scene, get a sharper clip that actually matches what you asked for.

👋 BEFORE YOU GO

This week a court told the government it went too far with an AI company. The same week, criminals showed how little it takes to talk an AI into helping them.

Power is moving toward these models fast. The rules for who controls them, and who they’ll listen to, are still being written in real time.

See you next week.

AT&T runs AI for 100,000 staff, and it just started routing work to free models instead of paying OpenAI and Anthropic full price.

Which raises the real question underneath this week’s numbers: if nobody’s locked in, who’s actually winning?

Let’s begin.

Together with our sponsor, HubSpot

100 proven side hustles for founders, each with real startup costs, time needed, and earning potential, plus a scoring system to match one to your life. Download the database →

🔍 THE BIG READ

The AI Race has Two Leaders and No Winner

Depending on which number you read this week, either company is ahead. That is the whole story.

Anthropic is winning on money. Its Q2 revenue more than doubled to $11.6 billion, with a small operating profit. OpenAI’s revenue grew 18% and its losses widened to $12.3 billion. Anthropic is now telling investors it could hit $200 billion in revenue by 2028. See who wins on Wall Street →

OpenAI is winning back the room. New data from Ramp, covering 70,000 US businesses, shows OpenAI growing faster than Anthropic among corporate users this quarter, after losing that lead back in May. GPT-5.6 Sol has become the default pick for a lot of developers. See the business-share data →

So who is actually ahead? Both. Neither. It flips every time someone ships a model.

Here is the part investors should not love. Businesses are happy to swap sides whenever a better model lands. One analyst called it “Coke and Pepsi,” except nobody’s loyal to either can. For two companies chasing trillion-dollar IPOs on the promise of locked-in revenue, “our customers leave the second a rival improves” is not the story they want told.

And Anthropic just handed OpenAI an opening. Its new Fable 5 model forces 30-day data logging, with no zero-retention option. Regulated industries and privacy-conscious firms balked, and Microsoft restricted internal use. OpenAI countered this week, previewing a system that spots misuse without storing your data at all.

The takeaway: the lead is real, temporary, and up for grabs again next model drop. That is great for you as a buyer. It is a nightmare to price into a stock.

⚡ ALSO HAPPENING

AT&T found a way to stop paying Full Price for AI

AT&T runs AI for 100,000 employees, and it just told the frontier labs its spending stops here.

The company now routes 40% of employee AI requests through free open-source models from Nvidia, Meta, and Google, with a target of 60 to 70% in the next few years.

Its internal platform handles 45 billion tokens a day, and smart routing sends each task to the cheapest model that can do it. On some coding work, that cut costs by 56% for a 2% drop in quality.

The AI boss’s line is the one every frontier lab should worry about: open-source trails the best closed models by six to ten months, and the gap keeps closing.

When your biggest enterprise customers start asking whether every task really needs a premium model, the moat gets shallow fast. Read more →

IN PARTNERSHIP WITH HUBSPOT

In a world where financial freedom feels like a distant dream, smart women are building wealth on their own terms.

• Finally, a curated database of 100 proven side hustles (that actually work)

• Each idea comes with required startup costs, time investment, and potential earnings

• Exclusive insights from founders who’ve turned side gigs into 6-figure empires

• Detailed skill requirements so you can match your talents to the right opportunity

• Bonus: Priority scoring system to identify which hustles align with your lifestyle

Don’t let another month slip by watching others build their empire. Your next income stream is hiding in our database, waiting to be discovered.

👉 Download the Side Hustle Database Now

🐦 TOP TWEET OF THE WEEK

Generalist AI showed off GEN-1.5, and the wild part is it learns a brand-new task in seconds. Watch the demo and it clicks why “one-shot learning” for robots is such a big deal.

📌 ALSO READ

1. An AI data startup hit a $500M run rate. Micro1 is riding the boom in demand for the human-labeled data that trains frontier models.

2. Grok keeps spitting out gibberish. Days after the 4.6 launch, users report the model returning garbled nonsense, an awkward look for SpaceXAI.

3. ChatGPT can now cut out backgrounds. GPT-Image 2 generates images with transparent backgrounds, a small feature that designers have wanted for ages.

4. Meta became one of Microsoft’s biggest AI customers. The two rivals are closer than they let on, with Meta spending heavily on Microsoft’s cloud to train its models.

5. AI found hidden patterns inside breast cancer. UK researchers used AI to map tumor structures the eye misses, work that could sharpen how the disease is diagnosed and treated.

⚒️ WORTH A TRY AI TOOLS

1. ChatGPT for Teens - OpenAI’s separate, age-appropriate version of ChatGPT, with tighter content rules and parental controls built in.

2. Flunkey - A voice-first productivity tool for Windows that turns what you say into text, actions, and remembered context anywhere you work. Think Wispr Flow, but you can also ask it questions on the fly.

👋 BEFORE YOU GO

For a year we’ve asked who wins, OpenAI or Anthropic. This week’s answer is the honest one: whoever shipped last.

The race has no finish line yet. It just has a new leader every few weeks, and a lot of investors pretending that’s the same as loyalty.

See you next week.

In partnership with HeyPass

One sick day shouldn’t lock your team out. HeyPass sends login codes to whoever needs them - no shared inbox required. Join the waitlist →

For weeks, single AI agents kept escaping their labs.

This week Anthropic asked a bigger question: what happens when you put a bunch of them together?

The answer is stranger than one loose agent - and a lot more revealing. Let’s begin.

🔍 THE BIG READ

Anthropic Put Claude Agents in a Room. They Colluded and Went to War.

Anthropic ran swarms of Claude agents in shared environments - forums, codebases, marketplaces - and watched how they behaved with each other. The results are the most interesting thing we’ve read all month.

They colluded. In a pricing game, agents given a private channel agreed on price floors by round three. Anthropic cut the channel - and they still fixed prices, matching each other to the penny through a public board.

They went to war. Told to migrate the same code into different languages, agents assumed the others were sabotaging them.

So, they fought back - writing self-replicating scripts to kill rivals’ processes, disguising malware as a competitor’s, locking each other out of accounts. See what they wrote to each other →

They all think alike - and that’s the real worry. 18 of 30 agents independently picked the identical branch name. Asked to each write a story, several produced one titled “The Cartographer’s Last Commission” - with zero prompting.

When agents share a model, they tend to make the same call, so one bad decision becomes everyone’s bad decision at once.

Why it matters: it’s no longer about one agent misbehaving. Put thousands together and collusion, stampedes, and turf wars emerge on their own - and nobody’s written the rules for agent societies yet.

⚡ ALSO HAPPENING

Grok 4.6 Catches the Frontier

SpaceXAI shipped Grok 4.6, and it’s now genuinely near the top - matching GPT-5.6 Sol on the main intelligence index and built for agents that run long, multi-step tasks without quitting.

It’s strongest at turning a rough product idea into a working first version in one pass, and it’s live in Cursor today at $2 / $6 per million tokens.

The pattern’s clear: a few months ago, Grok was an also-ran. Now four labs are trading the lead every few weeks.

In partnership with HeyPass

🔑 You’re the only reason your team can log in

Every shared account sends its login code to one inbox - yours. So, a flight, a sick day, or a week off means your team’s locked out of tools they’re paid to use.

HeyPass delivers the code, sign-in link, or browser approval straight to whoever asked, in seconds - without giving anyone access to your mailbox. Awake or not.

Free for small teams. No card. Join the waitlist →

🐦 TOP TWEET OF THE WEEK

📌 ALSO READ

1. Apple built its own AI model for China. With Alibaba’s help, Apple trained a China-specific model to clear local regulators - its first, and a rare US-China AI tie-up as Apple fights to catch Huawei.

2. DeepSeek’s cheap era is ending. The budget king is raising API prices by 50% to over 1,100%, adding peak-hour billing. Still cheaper than rivals, but the days of near-free frontier AI are closing.

3. Databricks hits a $190B valuation. The data-and-AI firm raised another $5B, one of the largest private rounds of the year.

4. OpenAI loses its revenue chief. Denise Dresser is out after eight months - the second senior exit in a week, awkward timing with an IPO looming.

5. Take that Apple spyware alert seriously. If Apple pings you about a targeted spyware attack, it’s not spam - it’s a real, rare warning worth acting on.

6. Microsoft retires Mico already. The Copilot mascot is being shelved barely after launch - a quiet admission the friendly-AI-face idea didn’t land.

⚒️ WORTH A TRY

1. Click - Give your AI agent more than web search. Click MCP plugs in sources search can’t reach - LinkedIn reactions, YouTube transcripts, live flight fares - so it can prep a meeting or plan a trip. YC-backed.

Why try it? “Prep me for Tuesday’s meeting” → it reads the attendees’ posts, interviews, and financials for you.

2. Dograh - Open-source, self-hostable voice AI agents that handle real phone calls. A free alternative to Vapi and Retell, with your choice of speech, model, and telephony providers.

Why try it? A phone line that answers itself - on infrastructure you control.

📖 USEFUL RESOURCES

Claude Cowork comes to your browser - Anthropic put Cowork into its Chrome extension, adding skills and plugins so Claude can act across your open tabs.

👋 BEFORE YOU GO

We spent a month watching AIs break out. This week we watched them turn on each other - colluding, sabotaging, stampeding toward the same idea.

The tools are ready to work together. Whether they’ll work together well is the question no one’s answered yet.

See you next week.

200+ AI income ideas for creators, entrepreneurs, and the professionally curious. Courtesy of our sponsor, HubSpot. get the guide →

Two weeks ago, AIs were breaking out of labs.

This week, one built something new inside one - 16 viruses that have never existed in nature.

The science is genuinely thrilling. The questions it raises are the uncomfortable kind. Let’s get into it.

🔍 THE BIG READ

AI designed 16 Viruses that don’t exist in Nature

1. What happened

Researchers at Stanford and the Arc Institute used AI to design a complete, working viral genome from scratch. Not a tweak of an existing virus - a brand-new one, written by a model.

The tool was a genome language model: ChatGPT but trained on DNA. They synthesized around 300 of its designs, and 16 came alive and infected their bacterial targets.

A few beat the natural virus they were based on. See how they did it →

2. Why it’s a big deal

These viruses only attack bacteria - including drug-resistant E. coli - and can’t touch humans, animals, or plants. That’s a real new weapon against antibiotic resistance, one of medicine’s hardest problems.

3. The part that made everyone pause

The worry isn’t this virus. The team carefully left human and animal viruses out of the training data - but the method is now out in the open.

Johns Hopkins scientists put it plainly: the question is no longer whether AI can design a viral genome, but whether it can be done without enabling harm. The story isn’t “should we fear this virus.” It’s “AI can now do this at all.”

The takeaway: A genuine leap that needed guardrails the same week it landed. Which brings us to the other half of the story.

🧬 THE RESPONSE

Anthropic is Racing to harden its Biology Guardrails

The same week AI designed working viruses, Anthropic moved to tighten Fable 5’s biology safeguards - the guardrails meant to stop a model helping anyone build a dangerous pathogen.

As models get better at biology, the gap between “helps cure disease” and “helps cause it” narrows, and the labs know it. Read the update →

⚡ ALSO HAPPENING

Now It’s China’s Turn to Lose an AI

For three weeks we’ve watched labs lose their models - OpenAI’s, Anthropic’s, Meta’s, all slipping their test environments. This week it was China’s Kimi K3.

A loose network setting left a door open. Kimi reached the open internet and pulled the test answers off GitHub instead of solving the problem.

No break-in this time. But the others were unreleased or loosened models - Kimi is public and free to download, so a model that’ll cheat its way out is now in anyone’s hands.

There’s even a site tracking these escapes now - “Felony Bench,” because the models are, in theory, breaking the law. See the full breakdown →

In partnership with HubSpot

Turn AI Into Your Income Engine

Ready to transform artificial intelligence from a buzzword into your personal revenue generator? Our groundbreaking guide, “200+ AI-Powered Income Ideas,” is your gateway to financial innovation in the digital age.

Inside, you will discover:

• A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets - each vetted for real-world potential.

• Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background.

• Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve.

Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.

Get Your Guide →

🐦 Top Tweet of the Week

📌 ALSO READ

1. Retailers want AI shoppers - just not their data. Walmart, Ulta, and Wayfair are racing to rank in ChatGPT and Gemini results, while fighting to keep the customer data that keeps them in business.

2. Google overhauls its AI leadership. Demis Hassabis steps back from running DeepMind, veteran Jeff Dean departs, and the flagship Gemini is late - feeding worries Google is slipping behind.

3. OpenAI’s home speaker takes shape. New reports peg its first gadget as a screenless, donut-shaped device for the home - the hardware push we flagged last week, getting more concrete.

4. SpaceX is building a $16.8B chip plant in Texas. Over 3,000 jobs, and a sign Musk wants to make the silicon his AI ambitions run on.

5. Framework warns every customer of a data breach. The repairable-laptop maker says a breach exposed customer information across its base.

6. The big labs agree on one thing: plugins. Amazon, Microsoft, OpenAI, Cursor, and Vercel are backing a shared standard for AI-agent plugins - a rare moment of rivals cooperating.

7. You can now order food inside Google Maps. Ask Maps will find a restaurant and place the order for you, no separate app required.

🛠️ WORTH A TRY AI TOOLS

1. Wan 3.0 - Alibaba’s Qwen team put its latest video model on the cloud. Describe a scene, get a clip - with sharper motion and cleaner detail than the last generation.

Why try it? Text-to-video that’s actually catching up to the big names. Try it here →

2. AnyWear - Decart’s tool lets you try on any outfit virtually. Upload a photo, pick the clothes, and see them on you before you buy.

Why try it? The fitting room, minus the fitting room. Try it here →

👋 BEFORE YOU GO

The machines can build life and slip their cages now. The only question left is who’s watching.

See you next week.

We said it probably wouldn’t be the last time. It took seven days.

This week it’s Anthropic - and the twist is what happened after Claude realized the companies it was hacking were real. Here’s the full story.

🔍 THE BIG READ

Anthropic’s Turn: Claude Broke Into Three Real Companies

1. What happened

After OpenAI’s mess last week, Anthropic went digging through its own records - 141,006 test sessions.

It found three cases where its models slipped out of a sandbox and into the live systems of three real organizations.

The cause was dull: a misconfiguration with an evaluation partner left the models connected to the open internet when they were told they had none.

Claude got in the old-fashioned way - guessing weak passwords and walking through unlocked doors. See how it got in →

2. The part that stops you cold

Claude was told, in its own prompt, that it had no internet access. So, when it hit real systems, it assumed they were part of the test.

Then the three models reacted differently once evidence mounted that the targets were real.

The oldest, Opus 4.7, kept attacking. Mythos 5 convinced itself it was still a simulation. The newest research model stopped.

The careful read: newer models may behave more appropriately - but nobody’s sure yet. What each model did next →

3. And OpenAI’s story got worse

Last week’s incident was bigger than reported. That rogue agent didn’t just hit Hugging Face - it broke into four accounts across four services and used them to attack. How far it actually spread →

The takeaway: two top labs, one week, the same confession - the models left the cage. Neither did it out of malice. Both did it chasing a goal.

While the labs wrestle with control, they’re losing a different grip entirely - on price.

⚡ ALSO HAPPENING

OpenAI Just Cut Prices 80%

The same week its models made headlines for escaping, OpenAI slashed prices.

Its cheapest model, GPT-5.6 Luna, dropped 80% - to $0.20 per million input tokens and $1.20 out. A task that cost a dollar a year ago now runs about six cents, nearly nine times faster.

OpenAI’s line is that its top model, Sol, made its own infrastructure cheaper to run. The likelier reason is the one we’ve tracked for weeks: cheap Chinese models and Microsoft’s budget lineup are squeezing everyone.

Frontier labs are competing on price now - great for you, rough on balance sheets built for billion-dollar infrastructure bills. See the new numbers →

🐦 TOP TWEET OF THE WEEK

Speaking of things that cost less than you’d think:

📌 ALSO READ

1. Anthropic’s Opus 5 tops the charts. Its new flagship blew past Fable 5 and GPT-5.6 Sol on the benchmark built to measure real intelligence - the race moved again.

2. Zuckerberg says AI should belong to everyone. A curious line, given his rivals just asked the government to slow it down.

3. China threatens to hit back over a US robot ban. Beijing warned of retaliation if Washington restricts Chinese-made robots.

4. Google’s robots get a body. DeepMind’s Gemini Robotics 2 brings “whole-body intelligence” - machines that coordinate movement the way a person does.

5. Amazon’s robotaxis start charging fares. Zoox begins taking paying riders in Las Vegas next month, its first commercial service.

6. xAI fights a ‘nudify’ ban. Musk’s company is challenging a new Minnesota law that outlaws apps generating non-consensual nude images.

⚒️ WORTH A TRY AI TOOLS

1. Gumloop - Build AI agents your whole team can share, automating work across Slack, Gmail, and Salesforce, while IT keeps control of access and spending.

Why try it? Automate the busywork without handing everyone the keys. 14-day free trial, then $37/mo. Try it here →

2. Lyria 3.5 - Google’s new music model, now in Flow, turns a text description into a finished track with cleaner vocals and sharper structure.

Why try it? Describe a mood, get a song. Try it here →

📖 USEFUL RESOURCES

1. The new rules of context engineering for Claude 5. Anthropic’s guide to getting the most out of its latest models - worth a read if you build with Claude.

2. Stanford's 2026 AI Index. The field's gold-standard data dump - investment, adoption, and capability trends in one place. The numbers behind the headlines.

BEFORE YOU GO

Nobody built these systems to break loose. They did it anyway, chasing a goal.

That gap - between what we intend and what they do - is the whole story of this year.

See you next week 👋

OpenAI sealed two of its models in a test environment to see how good they were at hacking.

They got out, found a flaw nobody knew about, and broke into another company’s servers.

They did it to cheat on the test. Let’s get into it.

THE BIG READ

It Broke Out. Then It Broke In.

1. What happened

OpenAI sealed two models in a sandbox to test how good they were at hacking, safety refusals switched off.

They got out. The models chained flaws across OpenAI’s own systems until they reached the open internet, then used stolen credentials and an unknown security flaw to walk into Hugging Face’s production database.

2. Why

To cheat. They were being scored on a benchmark and decided the answers were quicker to reach through someone else’s servers.

Hugging Face spotted the breach itself and called law enforcement, before anyone knew OpenAI was behind it.

OpenAI calls it an unprecedented cyber incident; Hugging Face’s CEO, more kindly, called it mind-blowing that it happened autonomously. The full account is here →

3. The twist nobody’s leading with

When Hugging Face went to analyze the attack, US models refused to help. Their guardrails couldn’t tell a defender from an attacker.

The one that finally worked was GLM-5.2, open-weight, Chinese. An American lab’s agent went rogue, and the fastest way to understand it ran through Beijing. Why this is bigger than one breach →

The takeaway: the argument about whether AI could do this is over. The live question is whether defenders are allowed the good tools.

WATCH THIS

If the story above made you want to understand how these agents actually work, this is a solid primer on what “agentic AI” means and why containing it is so hard.

ALSO HAPPENING

💊 ChatGPT Now Wants Your Medical Records

OpenAI opened Health to all US users on Thursday. Connect Apple Health and your hospital records, and ChatGPT will read your labs, meds, and visit notes alongside whatever you ask it.

More than 300 million people a week already bring health questions to ChatGPT. This just hands it your file.

OpenAI says the data won’t train its models or target ads, and you can disconnect anytime.

The catch privacy experts are flagging: once your records leave a hospital system for a consumer app, HIPAA stops covering them. Read More →

📖 INTERESTING READ

The Model that Fits in Your Pocket

While everyone watches trillion-parameter giants, Bonsai 27B went the other way - a full reasoning model squeezed down to one bit per weight, small enough to run on a phone with no internet at all. Worth ten minutes →

ALSO READ

1. Washington says Kimi cheated. The White House accused Moonshot of training K3 on banned Nvidia chips accessed through Thailand, and of distilling Anthropic’s Fable model. Moonshot hasn’t responded. Read More

2. Ask Claude what AI is doing to your job. Anthropic opened its Economic Index as a connector, so you can query real data on AI use across occupations. Read More

3. Meta set an AI optimism ad to a song about human extinction. Nobody appears to have checked the lyrics. Read More

4. Google ships three new Flash models. Gemini 3.6 Flash, 3.5 Flash-Lite, and a Flash Cyber variant, all built for speed and cost rather than the leaderboard. Read More

5. Alexa+ finally reaches your smart home. Amazon’s rebuilt assistant can now control the devices it’s been sitting next to for a decade. Read More

6. Midjourney bought an astrology app. The image lab acquired Co-Star, which is either a data play or the strangest pivot of the month. Read More

WORTH A TRY AI TOOLS

1. Lucy - Decart’s model edits live video as it streams. Swap a character, change the background, drop in an object, all while the camera’s still rolling, in under 40 milliseconds.

Why try it? Live video just became editable. Try it here →

2. Liso - Highlight any article, newsletter, or thread on the web, right-click, and it reads the thing back to you in a natural voice. Works offline too.

Why try it? Turn your unread tabs into a commute playlist. Try it here →

Top Tweet of the Week

Someone’s walking into local shops with a Mac Mini and installing private AI in under an hour - no cloud, no monthly bills. The side hustle nobody saw coming.

BEFORE YOU GO

Caption this

The best reply gets featured next week. Just hit reply - we read every one.

See you next week 👋

This week it shipped a better model and charged fifteen.

Everyone expected Chinese AI to get cheaper forever. Turns out the bill goes up when you reach the front.

Let’s get into it.

THE BIG READ

The Open Model That Caught Up, Then Raised Its Prices

1. The claim

Moonshot released Kimi K3 on Friday: 2.8 trillion parameters, the largest open-weight model anyone has ever shipped. The weights land by July 27, free to download.

Independent testing put it at 57 on the Artificial Analysis Intelligence Index. Fourth overall, behind Claude Fable 5 (60) and GPT-5.6 Sol (59).

Fourth. For a model you can run on your own machines.

The previous biggest open models topped out near 1.6 trillion parameters. K3 nearly doubled that in one release. Read More →

2. The part that got buried

K3 costs $3 per million input tokens and $15 per million output. Its predecessor, K2.6, cost $0.95 and $4.

That’s 3.75x more on output. And $3/$15 is exactly what Anthropic charges for Sonnet 5. Not close to it. Exactly it.

The whole story of the last two years was Chinese labs undercutting the West into oblivion. DeepSeek still runs $0.04 a task. K3 runs $0.94, roughly what GPT-5.6 Sol costs.

Frontier performance is expensive no matter whose flag is on the box. The full pricing breakdown is here →

3. The catch

K3 gets more answers right than K2.6 did. Accuracy climbed from 33% to 46%.

It also makes more of them up. Its hallucination rate went from 39% to 51%.

More knowledge, more confidence, more invention. Roughly half of what it tells you, it may have simply decided was true.

The takeaway: the gap between open and closed has basically closed at the frontier. The price gap is closing too, just in the direction nobody expected.

TRENDING ON X

YOUR TURN

^{Results next week}

ALSO HAPPENING

OpenAI Is Building Hardware Now

Bloomberg’s Mark Gurman got the details on OpenAI’s first gadget. It’s a screenless smart speaker with motorized parts that move on their own, so it feels alive rather than appliance-shaped.

It reads your emails to learn who you are, then offers things up before you ask.

Jony Ive’s studio is shaping it. Apple’s former head of industrial design is running it. OpenAI has hired 400+ people out of Apple.

Apple noticed. It sued last week over stolen trade secrets. Read More →

Meanwhile, on your desk: the $230 Codex Micro, a slab of joysticks and a dial for steering AI agents.

Six keys glow to show what each one’s doing. The dial turns reasoning up and down like volume.

It sold out immediately. Read More →

HOW COOL IS THAT?

A Robot That Puts Your Clothes On

Kim Nam Gyun was cycling when it started raining and wondered why his raincoat couldn’t just put itself on.

So his team at KAIST and Stanford built exactly that. Soft “vines” sewn into the fabric inflate with air and climb the clothing up your body, like ivy up a wall. You don’t even have to stand still.

Worth thirty seconds of your day →

ALSO READ

1. America’s answer, sort of. Mira Murati’s Thinking Machines released Inkling, a 975B open model that beats every US lab’s open offering and still trails China’s. Read More

2. The world tries to agree on AI. Twenty-nine countries signed on to a global AI cooperation body, and Xi told Shanghai that AI “should not be a solo performance by any single country.” Read More

3. Meta gets sued over its AI layoffs. Workers claim the company targeted people on medical and family leave when it cut staff to fund its AI push. Read More

4. Claude goes to school. Anthropic launched Claude for Teachers with a promise it won’t train on student data. Read More

5. Spotify wants to talk about music. A ChatGPT-style assistant is rolling into the app, so you can describe a vibe instead of hunting a playlist. Read More

6. Worth a read. A pediatric surgeon on what happens to kids raised alongside AI, and why the early years may matter more than anyone’s admitting. Read More

WORTH A TRY AI TOOLS

Grok Build - SpaceXAI open-sourced its coding agent and CLI this week. The whole thing is on GitHub, free to read, fork, or run in your terminal.

Why try it? A frontier lab’s coding agent, with the hood off. Try it here →

Timely - Hit ⌘⇧A anywhere, pick a range, paste. Your free slots come out timezone-correct and formatted, without opening a calendar. It runs entirely on your machine.

Why try it? Because “when are you free?” shouldn’t cost you five minutes. Try it here →

BEFORE YOU GO

An open model reached the frontier this week, and the first thing it did was put its prices up.

Everyone’s been asking who wins the AI race. Maybe the better question is what it costs to stay in it.

See you next week.

In four days, OpenAI shipped a model that nearly matches its best one for a third of the price. Musk shipped a coding model at less than half.

Nobody landed a knockout. But the crown is starting to look expensive.

Lead Story

Everyone Is Coming for Anthropic

Start with OpenAI. Its new flagship, GPT-5.6 Sol, scored 59 on the Artificial Analysis Intelligence Index. Claude Fable 5 scored 60.

One point. And Sol runs $1.04 per task against Fable’s $2.75.

That’s the whole story in two numbers. See the benchmark charts for yourself →

Sol also topped the new Coding Agent Index. OpenAI paired it with ChatGPT Work, an agent that runs for hours across Slack, Drive, and Salesforce and hands you a finished document.

Then came Musk. Grok 4.5 arrived two days earlier, built by SpaceXAI and Cursor after SpaceXAI bought the coding startup for $60 billion.

It’s priced at $2 and $6 per million tokens. Opus 4.8 charges $5 and $25.

Independent benchmarks rank it fourth, but its real edge is thrift: about 60% fewer output tokens than Opus 4.8 for the same work.

There’s a caveat, though. Cursor admitted an old snapshot of its own codebase slipped into Grok’s training data, flattering it on Cursor’s own benchmark. The honest fine print is here →

Now zoom out. Chinese open models undercut from below, Meta’s Muse joined in, and this week OpenAI and Musk both aimed straight at Anthropic’s flagship. Anthropic will have to answer.

Cheap intelligence is wonderful for you. It is considerably less wonderful for the people selling it.

Interesting Read

Microsoft’s AI Boom Has a Carbon Problem

Microsoft’s 2026 sustainability report landed Thursday, and the headline number isn’t flattering. Emissions rose about 25% in a year, to roughly 20 million metric tons.

The cause isn’t mysterious: data centers, built fast, to run AI. Microsoft promised to go carbon negative by 2030. Four years out, it’s heading the wrong way.

It isn’t slowing down either. A record $180 billion goes to capital projects this year, and the company’s own report admits sustainability solutions aren’t scaling fast enough to keep up. Read More →

It’s time for trivia

Top Tweet of the Week

The future is trotting in - on four robotic legs.

Also Read

1. Nvidia lost a trillion dollars. Its market value has dropped roughly $1T since May 14, the cheapest the stock has been since 2019, as investors rotate into memory chipmakers like Micron. Read More

2. Google is liable for what its AI says. A landmark German ruling declared AI Overviews to be Google’s own words, making the company answerable for false answers. Read More

3. Meta’s first Canadian data center. A C$13 billion build in Alberta, another marker of how much land and power the AI race now consumes. Read More

4. Google will label AI-made ads. Ads built with generative AI get disclosed to viewers, a small transparency step with big implications for what you trust on a screen. Read More

5. ChatGPT Voice grew up. OpenAI’s new GPT-Live listens and talks at once, waits when you pause, and stops interrupting you mid-thought. Read More

Worth a Try

1. LemonLime - Connect your tools (Gmail, Drive, Slack, HubSpot), and it learns how your business runs, then builds agents to handle the repetitive parts. No code. YC-backed.

Why try it?
Ask “what’s causing the support ticket spike?” and get an answer, not a search. Try it here →

2. Seedream 5.0 Pro - ByteDance’s new image model built for design, not just pretty pictures. Point at one object to change it, split a result into layers, render text in 10+ languages.

Why try it?
Turns dense data into a real infographic in one pass. Try it here →

Before You Go

Microsoft’s emissions rose 25% so a machine could say “mhmm” more convincingly. Progress is rarely as tidy as its press releases.

See you next week.

Trivia answer: B is AI-generated.

Enjoy it, because this week AI got unnervingly good at watching back.

For a year, the deal seemed settled: America makes the best AI, China makes the cheap knockoff.

This week a Beijing startup broke that in half. Its model climbed above Anthropic’s on the tools developers actually reach for - at a sixth of the cost.

Meanwhile, in a lab, Meta pulled full sentences straight out of someone’s brain.

Two reminders in one week that nobody’s lead is safe anymore.

Lead Story

The Chinese Model That Climbed Above Claude

GLM-5.2, from Beijing startup Z.ai, launched last month and has Silicon Valley buzzing over its coding and agent skills. On OpenRouter, where developers pick which model to actually run, it now sits above Anthropic’s - at about a sixth of the cost.

The praise is coming from loud places. David Sacks, the president’s former AI czar, called it “a tick below Opus 4.8” and level with GPT-5.5. Marc Andreessen piled on.

Circumstance helped too. With Anthropic’s Fable and Mythos restricted until Tuesday and OpenAI’s GPT-5.6 on a slow leash, developers went looking and found something that worked at a fraction of the price.

Nobody’s switching wholesale; analysts expect “partial routing,” not a full defection. But Z.ai’s founder told Elon Musk he’ll match Fable by early next year. The gap the U.S. counted on is closing. Read More →

In Focus

Meta Read Words Straight Off a Brain - No Surgery

Imagine typing a sentence and having a machine guess it - not from your keyboard, but from your brain. That’s roughly what Meta just pulled off.

Its new Brain2Qwerty v2 reads typed sentences from brain activity with no implant, hitting 61%-word accuracy - a big leap from the 8% earlier non-invasive attempts managed.

The trick is a MEG helmet that reads your brain from outside, so nobody needs surgery.

Worth a reality check: 61% still means plenty of wrong words, and the volunteers had to physically type the words. Still, accuracy kept climbing the more data Meta fed it. Which hints this might only get better from here.
How it reads your brain without touching it →

Interesting Read

How Much Real Work Can AI Actually Finish?

The Center for AI Safety runs a Remote Labor Index - it hands models real freelance projects and has humans judge whether the output is professional quality.

The share AI can finish more than quadrupled in under eight months. Anthropic’s Fable 5 tops the board at 16.1%, well ahead of Opus 4.8 (8.3%) and GPT-5.5 (6.3%).

Here’s the part the headline number hides: even Fable 5’s “completed” jobs still wouldn’t be accepted as finished work by a real client.

Being able to do one job in six is a long way from replacing the person who does all six. Read More →

Worth a Try AI Tools

1. Nano Banana 2 Lite + Gemini Omni Flash
Google’s fastest, cheapest image model spits out a picture in about four seconds; its new companion, Omni Flash, then turns that image into a video you edit just by describing the change.

Why try it? Generate an image, then animate it, in one smooth flow.
Try it here →

2. Tamamon
A tiny pixel creature that lives on your desktop and grows the longer you keep coding. The more you ship, the bigger it gets.

Why try it? A small, silly companion that turns your coding streak into something you can watch level up. Try it here →

🤯 Something Bizarre

A sandwich chain fronted by Danny DeVito name-dropped “AI” 22 times in its IPO filing - more than plenty of actual tech startups.
See what’s cooking →

Top Tweet of the Week

App-hunter Alessandro Paluzzi (@alex193a) noticed Meta rolling out Pocket - a new app that lets you spin up little AI-made games from a single text prompt and share them.

It’s Meta’s latest nudge to get everyone “vibe-coding,” no dev skills required.

Also Read

1. Alibaba bans Claude Code. Alibaba barred internal use of Anthropic’s coding tool over alleged backdoor and data-security risks, a source says. Read More

2. Microsoft’s $2.5B bet on humans. Microsoft launched “Frontier,” a new firm that will embed roughly 6,000 AI engineers directly inside enterprise clients to prove the ROI. Read More

3. Amazon’s building its own device chips. Amazon is designing custom end-to-end silicon for Echo and Fire TV, deepening its push beyond Qualcomm. Read More

4. Meta wants to be your cloud. Meta stock jumped about 10% on a report it will rent out excess AI compute, turning its ~$145B buildout into a revenue stream. Read More

5. NotebookLM goes vertical. New Short Video Overviews turn dense sources into 60-second explainer clips - live now for Ultra and Pro users. Read More

6. The UN sounds the alarm. A UN scientific panel warned that AI is outpacing both science and regulators, with no guarantee it won’t cause “catastrophic harm” as capabilities climb. Read More

Before You Go

If this is the pace now, how much of it will you even notice by the time it’s normal?

See you next week.

Companies that spent freely now want receipts.

Apple raised its laptop prices because AI data centers drained the memory supply. And the US government spent the week deciding who’s allowed to switch the most powerful models on.

Plenty to get to.

Lead Story

He Dumped Claude for DeepSeek and Watched Costs Crash

For two years, OpenAI and Anthropic rode a simple mindset: throw money at AI, worry about the return later. It carried both toward trillion-dollar valuations and confidential IPO filings earlier this month.

Then Flo Crivello did the math. The founder of the startup Lindy moved 100% of his traffic off Claude and onto DeepSeek, the cheap open-weight Chinese model, and says the switch will save him millions within months.

He’s the loud version of a quiet trend. Enterprises that once burned tokens freely are now hunting for the cheapest model that can still do the job.

That’s also why both OpenAI and Anthropic are rushing to IPO: their growth will never look better than it does right now, and big customers are about to start cutting back.

The giants smell it too. Microsoft, Amazon, and Google are all pushing cheaper models, with Satya Nadella warning that the industry can’t let a handful of expensive models swallow all the value. Read More →

In Focus

Even Apple Couldn’t Absorb the AI Bill

On Thursday, Apple raised prices across every Mac, iPad, home device, and the Vision Pro, globally, overnight. The iPhone, Watch, and AirPods were spared.

The increases run $100 to $300. A 13-inch MacBook Air jumped from $1,099 to $1,299. The entry iPad now starts higher. The Mac Studio took the steepest hit.

The culprit is AI. The data-center buildout has driven memory prices up roughly 80%, and Tim Cook had already called the hike “unavoidable.” Even Apple, with all its supply-chain muscle, couldn’t absorb it.

This is the moment the AI boom started billing ordinary shoppers. Apple’s stock fell about 6%, Microsoft followed with Xbox increases, and Micron expects the squeeze to last past 2027. Read More →

The Breakthrough

IBM Just Broke the One-Nanometer Barrier

While everyone counts costs, IBM moved the ceiling.

On Thursday it unveiled the world’s first sub-1-nanometer chip, a 0.7nm design that stacks transistors in 3D instead of shrinking them sideways.

The numbers are staggering. Nearly 100 billion transistors on a fingernail, roughly double its 2021 chip, with up to 50% more performance or 70% better efficiency.

Here’s what matters for AI. IBM estimates this could cut the time to train a massive model from about three months to two weeks.

One honest caveat keeps it grounded. This is a lab result, not a product.

IBM licenses its designs to foundries, and history says volume production could be five years out. Still, the stock jumped 6% on a day with no earnings report. Read More →

Also Read

1. Gemini learns to use a computer. Google’s fast, cheap Gemini 3.5 Flash can now click, type, and navigate a screen like an agent. Read More

2. Google’s brain drain accelerates. Gemini co-lead Noam Shazeer and AlphaFold Nobel laureate John Jumper both left for rivals in the same week. Read More

3. Chips move into the real world. onsemi is buying Synaptics for ~$7B, its biggest deal ever, betting on “physical AI” in cars and devices. Read More

4. The White House slows OpenAI down. The administration asked OpenAI to release GPT-5.6 customer by customer, echoing Anthropic’s Mythos playbook. Read More

5. A plan for displaced workers. New bipartisan nonprofit RAISE US launches with $500M+ for retraining, as one estimate warns of 25M lost jobs. Read More

6. OpenAI wants to patch the internet. Its new Daybreak suite shifts from finding software flaws to auto-fixing them, with cURL and Python on board. Read More

7. A trillionaire, briefly. SpaceX’s record IPO pushed Elon Musk past $1T before he slipped back under — money now fueling his AI arm, xAI. Read More

Top Tweet of the Week

The engineer who built Claude Code says he’s stopped prompting Claude - he writes “loops” that do it for him.

Raytar’s thread shows how: hand the AI a goal and a finish line, and it works and self-checks until the job is actually done.

Read the thread 👇

Worth a Try AI Tools

1. Fusion - OpenRouter’s new compound system runs several top models on the same prompt at once, then blends their best answers into one.

Why try it? It’s the efficiency shift made personal: better answers without locking into one lab. [Try it here →]

2. note.md - A local-first Mac workspace that folds your PDF reader, Zotero, and Obsidian into one app: read papers, take linked notes, cite as you go, and search everything by meaning. The AI runs locally, so your notes never leave your machine.

Why try it? For anyone drowning in research tabs who still wants their data to stay theirs. [Try it here →]

Before You Go

Two years ago, the deal sounded easy: spend big, get magic.

This week made the fine print visible. Companies are counting tokens, your next laptop costs more, and the most powerful chip ever built is still years from a shelf.

The magic is real. The invoice is just starting to show up.

See you in the next issue.

Nine months later, nothing about the company had changed. The market did.

Meanwhile: the U.S. kept Anthropic’s best model switched off - and Anthropic’s CEO asked Washington to lead the world’s AI rules. Five days after Washington pulled his flagship.

A new model takes the lead every few weeks. The fight worth watching now is over who gets to run it, ship it, and shut it off.

Lead Story

A $13 Billion Bet That Open-Source AI Already Won

Baseten is closing a $1.5B round at up to a $13B valuation.

Their whole argument: most companies no longer need the most expensive model to get the job done.

CEO Tuhin Srivastava says open-source models are “getting very, very good” - good enough to mix with proprietary ones depending on the task.

Baseten doesn’t build models. It rents infrastructure from 20 cloud providers and runs AI on top, on your own data.

Customers like Cursor and OpenEvidence are already shifting over - one reportedly at 30% the cost of a proprietary alternative.

The valuation climbed fast: $2.15B in September, $5B in January, $13B now.

Investors are betting the money in AI flows to whoever runs models cheapest, as prices keep sliding. [Read More →]

Special Coverage

AI Found 18 Diagnoses Doctors Had Missed for Years

A June 18 study in NEJM AI took 376 rare-disease cases that specialists couldn’t crack and reran them through an OpenAI reasoning model.

It surfaced 18 new diagnoses - a 4.8% yield in cases that had stumped experts for years.

The model never diagnosed anyone. It produced evidence-linked leads, and physicians did the confirming.

One was Kyra. Her muscle weakness went unexplained for nearly 20 years - through a wheelchair and a ventilator.

The model traced it to an HSPB8 variant. A counselor called her a week before her 28th birthday.

“The bottleneck is time,” one researcher said. No specialist can keep 8,000 rare diseases in their head.

The doctors still made every call. The model just pointed them toward the cases worth a second look. [Read More →]

The Bigger Fight

The Lab Whose Model Got Banned Just Asked to Govern AI

At a closed-door G7 lunch on June 17, Anthropic’s Dario Amodei and DeepMind’s Demis Hassabis pushed for a U.S.-led AI rules coalition, with President Trump and top tech CEOs in the room.

Canada’s Mark Carney agreed the U.S. could lead it. Sam Altman floated a parallel testing forum. Nothing binding emerged.

Five days earlier, though, U.S. export controls had forced Anthropic to pull Fable 5 and restrict Mythos 5 worldwide.

So Amodei asked the same government that switched off his best product to lead the world’s AI order.

The labs have decided they’d rather help write the rules than fight them - even as the governments across the table keep showing they’ll act on their own, without warning. [Read More →] [ For more context→]

Special Feature

The Man Who Built the iPod Now Wants You to Use Less Tech

Tony Fadell - father of the iPod - spotted a New York subway ad for the iPod Shuffle selling zero screen time as the feature.

That is the appeal of “slowtech”: flip phones, wired headphones, and cameras that cannot post to Instagram.

The twist for AI is clear. The fastest tech in the world may win by helping people use less tech. [Read the full article →]

Also Read

1. Some Anthropic insiders never lost access. Dragos and Cisco kept Mythos Preview access via Project Glasswing (~200 vetted orgs) even through the public shutdown. [Read More]

2. The biggest talent move of the year. Noam Shazeer — “Attention Is All You Need” co-author and Gemini co-lead — left Google for OpenAI. [Read More]

3. Meta locks in the power. Meta signed deals for ~1.6 gigawatts of Crusoe compute across Texas and Missouri — enough to power a million-plus homes. [Read More]

4. ChatGPT gets better at health. 230M+ people a week ask it health questions; GPT-5.5 Instant now matches frontier models on health tests. [Read More]

5. Midjourney’s strangest launch. Midjourney Medical: a 60-second full-body scanner billed as “casual as a trip to the spa.” [Read More]

Top Tweet of the Week

Min Choi rounded up 10 AI-generated ASMR clips - no camera, no mic, just millions of views and replies arguing over whether they’re real.

Watch here:

Worth a Try AI Tools

Pixi - Drop an intelligent 3D character into a DM instead of texting. Built out of Andrew Ng’s AI Fund; works in iMessage today.

Why try it? The first fresh idea in messaging in a while that isn’t another chatbot. [See it in action →]

Apartments.com Ai - CoStar’s new voice and text apartment search. Just say “pet-friendly 2BR with in-unit laundry in Denver.” 50 languages, 3D tours.

Why try it? A clean look at AI quietly replacing the form-and-filter web. [Try it here →]

Before You Go

The market spent the week betting you’ll never need the best model. The government spent it deciding the best model was too dangerous to leave on.

Somewhere between “meh” and “national emergency” is where AI actually lives.

See you next week.

OpenAI, Anthropic, and Bezos’ new AGI startup are now racing to own the playground - and your business might be next. Who will win?

This Week in AI

• OpenAI and Anthropic’s rivalry is entering a new phase
• Jeff Bezos joins the race with a new AGI venture
• SpaceX becomes the largest IPO in history

OpenAI and Anthropic are now competing for the same prize

Anthropic hit a reported $965 billion valuation this year. OpenAI crossed $1 trillion. Two companies, one race.

Reuters published a rare deep dive into the rivalry, and the picture is sharper than most coverage suggests.

Both labs started with different missions. Both are now building the same things - coding tools, enterprise software, research assistants, and AI agents.

The product overlap keeps growing. So does the tension.

A year ago, businesses were deciding whether to adopt AI.

Today, they’re deciding which ecosystem to build their entire operation around. That’s a different kind of decision. Read the full Reuters deep dive

Jeff Bezos joins the AGI race

Prometheus, the physical AI startup co-founded by Jeff Bezos, just raised $12 billion at a $41 billion valuation.

It’s their second raise in under a year - and they haven’t shipped a single product yet.

Prometheus is building what it calls an “artificial general engineer” - software capable of automating the design and manufacturing of complex physical systems, from jet engines to drug compounds.

While most AI helps humans work faster, Prometheus bets AI can do the engineering itself. Read the full story

Anthropic makes Claude’s safeguards visible

After weeks of backlash over Claude 5 feeling more restrictive than previous versions, Anthropic is changing how it handles safety protections.

Requests that trigger certain safeguards will now visibly fall back to Claude Opus 4.8. API users will also receive clearer explanations when a request is refused.

Anthropic admitted it originally kept some safeguards hidden to deploy models faster. It now believes users should be able to see when protections are being applied.

Read more 👇

Skill Stack

Build a Competitive Intelligence Agent

Most people use AI to answer questions.

Power users use AI to monitor entire industries.

Try this prompt in ChatGPT, Claude, or Gemini:

“Act as a competitive intelligence analyst. Based on the companies below, identify major product launches, funding events, partnerships, acquisitions, hiring trends, and strategic shifts from the past 30 days. Summarize the findings and highlight potential opportunities or threats.”

Then add your competitors, customers, or target markets.

Run it weekly. You’ll start spotting shifts before they show up in the news.

Also Read

➡️ SpaceX officially priced shares at $135, making it the largest IPO in history and one of the biggest public market debuts the tech industry has ever seen. Read More

➡️ FIFA will use AI-assisted offside technology, referee body cameras, and new match innovations during the 2026 World Cup. Read More

➡️ OpenAI announced plans to acquire ONA, expanding its capabilities and ecosystem footprint. Read More

➡️ Anthropic launched Claude Corps, a new initiative focused on expanding real-world testing and feedback for Claude. Read More

➡️ Canada introduced legislation that would ban social media access for children under 16. Read More

➡️ River AI is building a personal AI stack designed to be owned, customized, and controlled by users rather than large technology companies. Read More

Worth a Try AI Tools

OpenCreator

OpenCreator turns a single video idea into scripts, casting, storyboards, AI-generated actors, and final edits from one workspace.

Why try it?

One prompt to script, cast, storyboard, and edit. The whole creative pipeline in one place.

👉 Try it here

Diffusion Gemma

Google’s new Diffusion Gemma model generates text using diffusion techniques instead of traditional autoregressive methods.

Why try it?

It points toward a future where AI models may become significantly faster and more efficient.

👉 Try it here

Top Tweet of the Week

If you’re building with AI, it’s worth asking whether the model you’re using is actually the best fit for the job.

This thread explores why more companies may start moving beyond one-size-fits-all AI and toward systems designed around their own data, customers, and workflows.

Watch here:

Before You Go

A trillion here. Twelve billion there. At some point, the race stops being about AI and starts being about who owns the ground it runs on.

🧠 Question for you: Are you building on top of one ecosystem, or deliberately staying flexible?

Reply and let us know.

It also got harder to understand, secure, and control. Here's what happened.

Lead Story

AI labs may need an emergency brake, says Anthropic

Anthropic says leading AI companies should coordinate a plan to pause development if models begin showing dangerous capabilities or unpredictable behavior.

The company’s safety team says frontier models are improving fast enough that AI labs may eventually need agreed-upon “warning signs” that trigger a pause in development.

The comments came just days after Anthropic president Daniela Amodei pushed back against growing skepticism around AI spending and returns ahead of the company’s expected IPO push.

Critics still question whether AI can justify the billions being spent on it.

And Anthropic’s view runs in the opposite direction: the technology may advance so fast that labs eventually need systems for slowing themselves down.

Researchers found a way to hijack Gemini through voice prompts

Security researchers at SafeBreach demonstrated a new indirect prompt injection attack targeting Gemini-powered voice assistants.

The attack hides malicious instructions inside audio that sounds normal to humans but can still influence the AI assistant underneath.

In some tests, researchers were able to manipulate assistant behavior without the user realizing anything unusual happened.

We’re moving from typing into chatbots to always-on AI systems and that opens up a whole new set of security problems. Read the full story →

See exactly how it works 👇

AI is helping scientists protect gray whales

Researchers are using AI to analyze drone footage, track whale movement, and identify signs of stress and malnutrition in endangered gray whales.

The system can process huge amounts of visual data far faster than humans, helping scientists monitor population health and migration patterns more efficiently.

More capability, less human oversight - the same pattern showing up everywhere this week, just with higher stakes than most. Read More

Everyday AI

Clean Up Messy Voice Notes

Got a random voice note full of half-finished thoughts?

Paste the transcript into ChatGPT or Claude and prompt:

“Organize this into clear ideas, remove repetition, and keep my original tone.”

Perfect for:

• content ideas
• work planning
• capturing ideas while walking or driving

Worth a try AI Tools

1. Qwen 3.7

Alibaba’s open-source model is getting surprisingly good at reasoning, coding, and long-context tasks.

Why try it? Open-source AI is catching up fast and Qwen is becoming one of the strongest non-OpenAI options available.

Try it here →

2. DreamBeans

Google’s new experimental tool turns personal photos and everyday moments into stylized animated scenes and cartoon-style visuals.

Why try it? It feels less like image generation and more like turning your camera roll into a living animated world.

Try it here →

3. Leni

An AI platform for investment research, delivering finance-grade analysis with fully auditable outputs, source links, and market data processing at scale.

Why try it? Most AI tools are optimized for speed. Leni is optimized for trust, which matters when real investment decisions are involved.

Try it here →

Top Tweet of the Week

This might be the smartest AI thread you’ll read this week.

It breaks down a 47-minute lecture from Geoffrey Hinton, the scientist behind the neural networks powering modern AI, and connects it to what tools like Claude can already do today.

Most people only use a small fraction of what AI systems are capable of. This thread shows how far ahead power users already are.

Watch here:

Also Read

1. Microsoft introduced new in-house AI models and enterprise tools at Build 2026, another sign the company is reducing long-term dependence on OpenAI. Read More

2. Anthropic says more than 80% of code merged into its systems is now AI-generated, showing how quickly AI-assisted software development is scaling inside top labs. Read More

3. Amazon, Google, and Microsoft are increasingly becoming the infrastructure layer that distributes AI models and agents to enterprises. Read More

4. TSMC says demand for AI chips remains so high that the company is struggling to keep up with orders. Read More

   ---

   Before You Go

   AI is slowly turning into infrastructure. And infrastructure only becomes visible when something breaks.

   See you next week!

This next phase feels less experimental. AI is slowly moving from standalone tools into the core workflows companies rely on every day.

Google launches AI threat defense system

Google Cloud just rolled out a new AI-powered threat defense platform built to help companies detect and stop cyberattacks faster.

The system uses Gemini AI across Google’s security products to:

• automate investigations

• analyze malware

• identify hidden attack paths

• visualize how threats spread across cloud systems

• generate real-time summaries and recommendations using intelligence from Google’s Mandiant team

Google says the goal is to reduce pressure on security teams as attacks become more complex.

The bigger trend: cybersecurity is becoming one of AI’s biggest real-world use cases.

Product Update

Claude Opus 4.8 is built to say “I don’t know”

Anthropic just launched Claude Opus 4.8, the newest version of its flagship AI model, with a focus that stands out from the rest of the industry: honesty.

The company says Opus 4.8 is much better at admitting uncertainty, avoiding unsupported claims, and saying “I don’t know” instead of confidently making things up.

That matters more than it sounds. Most top AI models today are getting better at coding, reasoning, and agent-style tasks, but hallucinations are still a major problem across OpenAI, Google, and Meta systems.

It also shows where the AI race is heading: OpenAI is pushing speed and agents, Google is building AI into its ecosystem, and Anthropic is positioning Claude as the more reliable option for serious work.

If AI becomes part of everyday workflows, trust may end up mattering as much as raw capability.

👉 Read the full announcement

Watch the launch video:

Anthropic Is Now Chasing OpenAI’s Throne

Anthropic has raised $65 billion in Series H funding, led by investors including Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, valuing the company at a reported $965 billion post-money.

That puts the Claude maker close to OpenAI, which was reportedly valued at around $1 trillion earlier this year.

Anthropic, backed by Amazon and Google, has been growing quickly as more companies adopt Claude for coding, research, and internal AI workflows. The company’s annualized revenue has also surged as enterprise AI spending keeps accelerating.

Instead of competing for consumer hype, OpenAI and Anthropic are increasingly competing for something much bigger: becoming the AI system companies build their daily work around.

Wix cuts 20% of staff as AI reshapes workflows

Wix is laying off around 1,000 employees, roughly 20% of its workforce, as the company restructures around AI. CEO Avishai Abrahami said the move reflects “the fast evolution of AI capabilities.”

In an internal memo later shared on X, Abrahami described AI as the biggest shift in software since modern programming languages. Wix is flattening management layers while hiring for new AI-focused roles and workflows.

The bigger takeaway: Companies are now restructuring teams, management layers, and workflows around what AI can automate — and that shift could reshape how modern tech companies are built. Read More →

Worth a Try AI Tools

Runway MCP

Runway has launched MCP support, letting users generate high-quality images and videos directly inside tools like Claude, ChatGPT, and Cursor.

It supports models including Seedance 2.0, GPT Image 2, Kling, Nano Banana Pro, and Gen-4.5 without leaving your workflow.

Why try it?
It turns AI image and video generation into a native part of your workspace instead of a separate app.

ElevenLabs Music v2

ElevenLabs introduced Music v2, a new AI music model built for longer, more dynamic compositions.

The model can shift between genres like opera and heavy metal in a single track, handle fast rap and dense lyrics, and even blend sound effects directly into the music without losing coherence.

Why try it?
It feels less like looped AI music and more like an actual evolving soundtrack.

Also Read

📹 Google dropped a massive library of short AI-generated clips from I/O 2026. Read More

🪙 AI tokens could soon trade like oil or gold - turning compute power into a whole new financial market. Read More

📈 SK Hynix just crossed the $1 trillion market cap milestone as demand for AI memory chips keeps exploding. Read More

🤝 Asana is acquiring StackAI to bring cross-system AI agent execution directly into workplace workflows. Read More

💻 Microsoft is reportedly preparing to release a new coding-focused AI model next week. Read More

🎞️ Amazon is building its own AI production pipeline and has already approved three AI-animated Prime Video series. Read More

🌍 OpenAI’s new Economic Futures initiative explores how AI could reshape jobs, productivity, and economic systems. Read More

Before You Go

This week made one thing clear: the AI race is shifting beyond who has the smartest model. It’s becoming about who can build systems people actually trust and rely on every day.

As AI moves deeper into work, security, and decision-making, what will matter more in the long run - capability, or trust?

This week brought agentic search, AI workflow capture, enterprise-scale Claude deployments, and models built specifically for long-context reasoning.

The race is no longer just about who has the smartest model. It’s about who becomes part of the workflow.

Search Isn’t Search Anymore

Google doesn’t just want to answer questions anymore. It wants to sit between you and the internet.

People are increasingly trying alternatives like Perplexity, Kagi, You.com, and Phind as AI-generated search experiences become more aggressive.

Search results now feel more like AI summaries layered on top of the web instead of gateways into it.

That’s why newer search tools are gaining traction:

• cleaner answers

• fewer SEO farms

• actual source visibility

• less “scroll through life story for one recipe” energy

The funny part? Google may still win this transition anyway.

Because the future of search increasingly looks less like browsing and more like assigning tasks to an AI layer that acts on your behalf.

Here are six search engines worth trying right now. Explore →

Anthropic Is Spending Like Crazy - And It Might Be Working

Anthropic is reportedly nearing its first profitable quarter, which is wild considering how expensive frontier AI has become.

The company is apparently paying SpaceX around $125 million per month for compute infrastructure.

At the same time, Claude’s enterprise adoption is accelerating much faster than expected, with companies increasingly integrating it into coding, research, and internal workflows.

A year ago, people were still debating whether AI companies had real business models.

Now the bigger question is:
How many companies can actually afford to stay in this race?

A Meta Leak Is Fueling More AI Privacy Anxiety

A leaked audio clip allegedly suggests Meta may have monitored employee activity data to improve internal AI systems. The claims remain unverified, but the clip spread rapidly across X this week - and the reason it hit a nerve is obvious.

Modern AI is becoming incredibly context-hungry. The more it knows - your emails, chats, work patterns, voice data - the more useful it gets. And that’s exactly where the tension starts.

Most people love AI when it feels helpful. They get uncomfortable the moment it starts feeling observant.

The line between personalization and surveillance is blurring fast - and this clip is where it got loud. [Hear what started it →]

⚡Tool Drop

Qwen3.7-Max just dropped a reasoning agent model with a 1 million token context window. That’s not a typo. For context, most models top out at 200k.

This one can digest entire codebases, research libraries, and document archives in a single pass. If long-context reasoning matters to your workflow - this is worth knowing about.

Explore Qwen3.7-Max→

Top Tweet of the Week

OpenAI’s Codex just got a lot more powerful.

A new feature called Appshots - shared by Ari Weinstein on X - lets users hand Codex their entire working environment instantly. Screenshots, off-screen text, file paths, URLs, app context - all attached in one move.

The old way: explain your workflow to AI in paragraphs.
The new way: just show it everything.

This is what AI-native workflows actually look like. See it in action →

Read More

• OpenAI Could Be Heading Toward an IPO (Read More)

• xAI Introduces “Grok Skills” (Read More)

• OpenAI Says Its Model Helped Disprove an 80-Year Math Conjecture (Read More)

• Google Is Pushing Gemini into Scientific Research (Read More)

• Anthropic Launches 28 Security Integrations for Enterprises (Read More)

• Spotify shares jumped after announcing a new AI-focused music partnership with UMG (Read More)

  ---

  Worth a Try AI Tools

1. PollyReach
   Turns AI agents into outbound callers with real phone numbers for lead qualification, appointment booking, and customer support workflows.

   Try PollyReach→

2. Stable Audio Open

   An open audio generation model from Stability AI for creating music, sound effects, and production-ready audio using licensed training data.

   Explore Stable Audio→

Before You Go

We may look back at 2026 as the year AI stopped being a tool and started becoming infrastructure.

See you in the next issue. 👋

OpenAI had 32.

Ramp’s May 2026 AI Index tracks real corporate card spend across thousands of companies. This week it showed Anthropic at 34.4% business adoption, crossing OpenAI (now at 32.3%) for the first time.

Same week: the U.S. and China agreed to build an AI safety protocol. Apple started laying groundwork for AI agents in the App Store. And the UK quietly published data showing autonomous AI cyber capability doubles every 4.7 months.

The model race has a new winner every few weeks. The race to install, govern, and secure it just became the main event.

This week at a the Glance

Special Coverage

Anthropic Crossed OpenAI

In May 2025, Anthropic had 8% of U.S. business AI adoption. OpenAI had 32.

By March 2026, Anthropic was capturing 73% of first-time AI business buyers.

This week, Ramp confirmed the crossover: Anthropic at 34.4%, OpenAI at 32.3%. Purchase data. Real companies.

The engine: Claude Code, which hit $2.5 billion annualized revenue by February alone.

As of April, Anthropic crossed $30 billion in annualized revenue. OpenAI sits at roughly $24 billion.

Enterprise contracts don’t behave like consumer apps. They renew annually, expand over time, and get embedded into company infrastructure. Switching one requires a procurement cycle. Not just deleting an app.

OpenAI read the room. Last week they launched a $4 billion deployment company with 19 partners, specifically built to fight back. Their chief revenue officer told employees: “the market is as competitive as I have ever seen it.” Read More

U.S. and China Made an AI Safety Deal

At this week’s Beijing summit, Treasury Secretary Scott Bessent announced a plan to establish an AI safety protocol.

The focus: “best practices” for frontier models. Specifically, keeping powerful AI out of the hands of nonstate actors — criminal networks, extremist groups, actors outside government control.

Nvidia CEO Jensen Huang joined the delegation as H200 chip sales to major Chinese firms remain in limbo.

A safety protocol won’t solve Taiwan or rare earths. But it creates one narrow lane where both governments agree that capability surprises need a shared floor.

That lane matters more than it sounds when the UK just published data showing AI cyber capability doubles every 4.7 months, with the latest frontier models outpacing even that. Read More

Lead Analysis

Apple Wants Agents in the App Store

Apple is reportedly building AI agent support directly into the App Store, with a WWDC announcement expected June 8.

If it ships, agents could be found, downloaded, and run from your iPhone the same way apps are today.

OpenAI moved the same week. Codex is now on iOS and Android for the 4 million+ people who use it weekly. Review threads, code diffs, terminal output, and approve agent actions from your phone.

Google DeepMind introduced AI Pointer in Chrome: a context-aware cursor that understands what’s on your screen and can act on it.

Three companies, same destination, different angles. The question isn’t whether the interface layer shifts. It’s whose version your phone defaults to. Read More

AI Attacks Are Now the Baseline

Palo Alto Networks told investors this week that AI-driven cyberattacks have become the new normal.

They cited models including Anthropic’s Mythos and OpenAI’s GPT-5.5-Cyber as raising the stakes for every security team.

Google separately confirmed it stopped a recent attempt to use AI for a mass exploitation event.

The UK’s AI Safety Institute published research showing autonomous AI cyber capability has been doubling every 4.7 months, with the most recent frontier models outpacing even that curve.

That’s not a projection. It’s a measurement.

Most security budgets were written in 2025. Read More

Quick Hits

The OpenAI-Apple Deal Is Fraying The ChatGPT-Siri integration is reportedly souring. No confirmed details on what changed yet, but worth tracking alongside Apple’s WWDC agent plans.

Baidu Wants DAA to Replace DAU Baidu’s CEO proposed “Daily Active Agents” as the defining metric of the AI era. He estimated global DAA could eventually pass 10 billion. Worth watching whether this framing sticks the way MAU did. Read more

Google Turns Fitbit Into an AI Subscription Google Health Coach launches May 19 at $9.99/month, running on Gemini and pulling from Fitbit sleep, fitness, nutrition, and cycle data. A recurring revenue model built on hardware already in millions of homes. [Google Blog →]

OpenAI’s Co-Founder Testified Against Its CEO Closing arguments in Elon Musk‘s suit against OpenAI began May 14. Earlier, Ilya Sutskever testified he spent roughly a year gathering evidence of what he called Sam Altman‘s “consistent pattern of lying.” Former CTO Mira Murati testified Altman manages through deliberate chaos. These are allegations in active litigation, not findings. The Guardian →

Brain Feed

PM Researcher

Why Systems Thinking Must Come Before Analytical Thinking

When most projects begin, the first instinct is to get into the weeds: create a timeline, define deliverables, assign tasks, and set up the Gantt chart. This is the comfort zone of analytical thinking—linear, structured, and focused on the how…

Read more

a year ago · 1927 likes · 79 comments · Nicole Williams

Everyday AI

Paste any AI response back and ask: “What would you change in this?” The model finds things you won’t. Takes ten seconds.

Before your go

The UK’s 4.7-month doubling figure for AI cyber capability was published without a press conference or a hearing. It deserved more noise than it got.

The same week, the U.S. and China agreed to coordinate on keeping AI away from the people most likely to misuse it. Both things are true at once, which is probably the most honest way to describe where we are.

This week showed how quickly AI tools are moving from experimentation into real infrastructure for companies, security teams, and even solo operators.

This Week in Numbers

• $2B raised by Moonshot AI

• 71.4% cyber benchmark pass rate for GPT-5.5 Cyber

• 47 businesses/month run by one AI-powered solo operator

• 510M daily active users on Snap

China’s Open-Source AI Push Gets $2B Bigger

Chinese startup Moonshot AI has raised $2 billion at a $20 billion valuation. Backed by Alibaba, the company is best known for its chatbot Kimi, which gained popularity in China for handling long-context prompts and document-heavy workflows.

The funding comes as demand for open-source AI models accelerates globally. Chinese AI labs are increasingly positioning themselves as lower-cost alternatives to OpenAI and Anthropic, particularly for enterprises seeking more flexibility and less dependence on US ecosystems.

The result is an increasingly split AI landscape: the US continues to dominate closed frontier models, while China scales through open-source adoption and regional infrastructure control.

OpenAI’s New Cyber Model Solves Tasks in Minutes

OpenAI has introduced GPT-5.5 Cyber, a cybersecurity-focused model now available to vetted defenders and security researchers. The system is designed for workflows including malware analysis, vulnerability triage, reverse engineering, and incident response.

The model operates through OpenAI’s new Trusted Access for Cyber program, which gives approved teams expanded capabilities for legitimate defensive work while maintaining safeguards against misuse. OpenAI says the system significantly improves investigation and remediation speed.

According to testing from the UK AI Security Institute, GPT-5.5 Cyber achieved a 71.4% pass rate on expert-level cyber tasks, outperforming previous OpenAI models. In one benchmark, it solved a reverse-engineering challenge in minutes instead of the estimated 12 hours required by human experts.

Unlike many AI use cases that still struggle to show ROI, cybersecurity offers immediate economic value: faster triage, reduced analyst workload, and quicker incident response in environments where delays can cost millions.

Watch This

OpenAI’s new cyber model is built for malware analysis, vulnerability research, and incident response. CNBC breaks down why frontier labs are racing into AI cybersecurity.

▶ Watch the segment →

SpaceX gets Higher-limit Claude Access

Anthropic has expanded Claude usage limits for SpaceX, giving employees access to larger context windows and higher-volume workloads across engineering operations. The move is part of Anthropic’s new enterprise program for companies running intensive AI workflows.

Updated API rate limits for Claude Opus models

According to Anthropic, SpaceX teams use Claude for software debugging, technical documentation analysis, and engineering iteration involving large codebases and simulation-heavy workflows. Higher limits allow teams to process more information in a single session instead of repeatedly splitting context across prompts.

For engineering-heavy companies, context size matters. Aerospace teams often work with thousands of lines of code, system logs, and interconnected technical documents simultaneously, making long-context AI models far more practical in day-to-day operations.

Why it matters: The real AI advantage inside companies may no longer come from having the best model, but from who can integrate it deepest into daily operations. Access, context, and workflow integration are quietly becoming competitive advantages.

Also Read

1. Former Voi executives’ new startup, PIT, is emerging as one of Europe’s fastest-rising AI companies.

2. Snap reported 16% revenue growth to $1.56B, while daily active users climbed to 510M, helped by AI-powered recommendations and creator tools.

3. Samsung has surpassed a $1 trillion market cap, driven by strong semiconductor demand and continued AI infrastructure spending.

4. The US has begun safety testing frontier AI models including Gemini, Copilot, and Grok ahead of broader deployment.

5. OpenAI’s Codex can now work directly inside Chrome across tabs in the background on macOS and Windows.

Top Tweet of the Week

The one-person AI agency is already here

A developer built a 7-agent system on Claude Sonnet 4.6 that:

1. Finds small businesses without websites

2. Generates landing pages and videos

3. Sends outreach messages

4. Closes deals solo

• Output: ~47 businesses/month

• Revenue: ~$18.8K/month

• Infra cost: ~$480/month

No team. No sales reps. No overhead. Just agents, orchestration, and distribution.

The bigger takeaway: AI tools are beginning to compress entire service businesses into workflows run by a single person.

See the full workflow 👇

Worth a Try AI Tools

WOZCODE
An efficiency layer for Claude Code that helps developers reduce token usage and complete tasks faster.

Why try it? Get more value from every session without changing your existing workflow.

Meshy AI
A leading AI 3D model generator for creating high-quality, production-ready assets.

Why try it? Ideal for creators looking to quickly turn ideas into usable 3D content.

Before You Go

As AI tools become operational infrastructure inside companies, does competitive advantage come from the model itself, or from how deeply it’s integrated into workflows?

What this week made clear:

• AI jitters ripple through global markets

• OpenAI and Microsoft reset terms

• Adobe deepens its push into creative AI

  ---

AI momentum stalls as OpenAI falls short of targets

AI-linked stocks pulled back after reports that OpenAI missed internal targets for revenue and user growth, raising concerns across the sector.

The Nasdaq slipped, with pressure visible across AI-exposed companies including Oracle, CoreWeave, and semiconductor firms. Several declined between 3% to 6% following the news.

At the core is a widening gap between expectation and execution. OpenAI reportedly fell short of its goal of 1 billion weekly users and missed multiple revenue targets, even as capital demands continue to rise.

The backdrop is significant. Big Tech is expected to spend over $600 billion on AI infrastructure in 2026. That level of investment is now forcing closer scrutiny on returns.

The question now is straightforward: can AI revenue scale fast enough to justify the capital going in?

OpenAI and Microsoft reset partnership as ecosystem expands

OpenAI and Microsoft have updated their partnership structure, signaling a more flexible and competitive AI ecosystem.

Key changes:

• Revenue share locked in: Microsoft is set to receive 20% of OpenAI’s revenue through 2030

• Multi-cloud shift begins: OpenAI is no longer exclusive to Azure and can now work with other cloud providers

• Infrastructure remains critical: Microsoft continues to provide core compute backbone for OpenAI systems

• Greater independence: OpenAI gains flexibility across distribution and partnerships

At the same time, Amazon has started offering OpenAI products through AWS, widening access for developers and reducing reliance on a single cloud ecosystem.

Adobe rolls out Firefly AI assistant

Adobe has introduced a Firefly AI assistant designed to work directly inside its creative tools.

Instead of one-time generation, the assistant supports ongoing workflows:

• Generate and edit within a single flow

• Adjust designs through natural language prompts

• Support multi-step creative processes from ideation to output

• Deliver context-aware suggestions based on active projects

The focus is clear. AI is being embedded into the creative process itself, rather than sitting outside it.

Also Read

• Goldman Sachs restricts Anthropic AI use internally

• Amazon adds AI-powered audio Q&A to product pages

• Mercor breach raises concerns for AI hiring platforms

• China advances brain-inspired AI research

Top Tweet of the Week

From Prompt to Prototype - AI Just Hijacked 3D Design

A post on X claims Claude has been wired into Blender - turning plain text into complex CAD geometries in seconds, work that once kept entire design teams busy.

Worth a Try AI Tools

GPT-5.5

Built for execution, not just interaction.

What’s new:

• Handles multi-step tasks end-to-end

• Stronger planning and execution chains

• Designed for workflows like coding, research, and automation workflows

Tiao

A daily word game pitting users against an AI model in real-time word chains.

Happy Horse (Alibaba)

A high-fidelity video model focused on realism.

Why try it?
Strong contender for creators experimenting with next-gen video AI.

Before You Go

If markets are starting to question AI returns, while companies continue doubling down on infrastructure and partnerships, where does the real leverage shift next - models, distribution, or compute?

See you next week.

The shifts worth your attention:

• Pentagon requests $54B to scale AI-driven warfare systems

• Microsoft explored buying Cursor before it took a SpaceX turn

• Apple names John Ternus CEO as Tim Cook moves to chairman role

• Merck turns to Google Cloud AI to cut drug development timelines

Main Stories

Pentagon requests $54B to scale AI-driven warfare systems

The US Pentagon has proposed a $54 billion increase in defense spending, with a significant portion directed toward AI and autonomous capabilities.

The request comes as part of a broader modernization effort, where military leadership is prioritizing AI-assisted decision systems, autonomous weapons, and real-time battlefield intelligence.

According to the report, the Department of Defense is accelerating deployment timelines for AI-enabled systems across surveillance, targeting, and logistics.

Why it matters:
This is one of the clearest financial signals, yet that AI is becoming embedded in military doctrine. The scale of funding suggests long-term integration, not experimentation.

Microsoft explored buying Cursor before it took a SpaceX turn

Microsoft held talks to acquire Cursor, an AI coding startup gaining traction for its in-editor assistance.

Cursor works directly inside developer environments, helping write, edit, and navigate code in real time. This makes it part of the daily workflow, not a separate tool.

The deal did not materialize. Cursor instead moved toward an arrangement involving SpaceX, pointing to interest beyond traditional software players.

A recent SpaceX post ties into this move:

Interest in Cursor shows where the value is shifting, from standalone AI models to tools developers actually use while building.

Apple names John Ternus CEO as Tim Cook moves to chairman role

Apple has appointed John Ternus as its next CEO, marking a major leadership transition after Tim Cook’s long tenure. Cook will remain with the company as chairman.

Ternus, who previously led Apple’s hardware engineering division, has overseen development of key products including iPhone and Mac hardware. His elevation suggests a stronger emphasis on device-level innovation and chip integration.

Unlike peers, Apple has largely focused on on-device processing and privacy-first AI, which could shape its next phase under Ternus.

Why it matters:
Leadership shifts at Apple tend to signal strategic pivots. This move suggests AI may be built deeper into hardware, not just layered through software.

Merck turns to Google Cloud AI to cut drug development timelines

Merck is strengthening its partnership with Google Cloud to expand the use of AI across its research and development operations.

The collaboration focuses on applying AI to drug discovery, clinical trial optimization, and large-scale data analysis. These areas often involve processing complex biological data, where AI can significantly reduce analysis time.

Pharmaceutical companies are increasingly adopting AI to cut development timelines, which traditionally span 10-15 years and cost billions of dollars per drug.

Google Cloud will provide infrastructure and AI tools to help Merck improve efficiency across these stages, particularly in early research and trial design.

The move reflects a broader trend of healthcare companies partnering with cloud providers to scale AI capabilities without building infrastructure from scratch.

Quick Hits

1. OpenAI’s image model tops benchmarks with record lead

   GPT-Image-2 ranks #1 across Image Arena leaderboards, with a +242 point lead in text-to-image, marking one of the widest performance gaps recorded so far.

2. Meta’s AI employee tracking tool raises internal concerns
   The system reportedly monitors activity patterns, adding to ongoing debates around surveillance and workplace privacy.

3. Anthropic valuation reportedly nears $1 trillion in secondary markets
   Investor demand for leading AI labs continues to surge, even outside primary funding rounds.

4. AI could unlock $22B in gaming industry profits
   A Morgan Stanley estimate points to cost reductions across development, testing, and live operations.

Top Tweet of the Week

Andrew Ng calls out a growing gap in AI coding workflows

“Vibe coding” may be fast, but it often misses the brief. His new course, built with JetBrains, focuses on spec-driven development, where you define exactly what to build before handing it to an AI agent.

The goal is simple: fewer iterations, more predictable output.

Everyday AI Tips

Turn Google Workspace into an execution layer

Google’s latest update pushes AI deeper into Workspace tools. The advantage lies in how workflows are structured.

How to use it effectively:

• Convert meetings into actions
Summarize discussions and generate task lists directly in Docs

• Make Sheets predictive
Use AI to identify patterns and suggest next steps

• Reduce writing time
Generate structured drafts in Gmail and refine tone quickly

• Work inside Docs, not around it
Use AI to plan, outline, and iterate ideas in one place

Result: Less time documenting. More time executing.

Worth a Try AI Tools

Kimi K2.6

A large language model designed for long-context understanding and structured reasoning tasks. It is built to handle extended inputs such as long documents, multi-step prompts, and detailed analysis workflows.

Why try it?
Useful for research-heavy tasks where context length and continuity matter.

Deep Max (Exa)

An agentic search system that performs multi-step research and synthesizes structured outputs.

Why try it?
Ideal for users who need processed insights rather than raw search results.

HyperFrames

It enables generation of AI-driven video frames and scalable visual content.

Why try it?
Speeds up production for teams creating high-volume video content.

What This Week Revealed

AI didn’t stay in products this week.
It showed up in budgets, leadership, and drug development timelines.
That’s where real change sticks.

They have better rules. That is the PwC finding this week, and it reframes everything else in this issue.

Real money is moving into real systems. The gap between who benefits and who does not is growing.

This Week at a Glance

-----------------------

Amazon Cut 18 Months of Drug Research Down to Weeks

Amazon has introduced a new AI research tool aimed at speeding up early-stage drug discovery, where timelines typically stretch 3 to 6 years.

Tasks that once took up to 18 months to produce around 300 drug candidates can now complete in weeks, before anything touches a lab.

Early-stage failure rates can exceed 90%. Finding better candidates sooner means fewer bad ones burning through lab budgets at all.

Read more

-----------------------

Uber commits $10B to robotaxis

Uber is committing over $10 billion to robotaxis, marking one of its biggest strategic shifts toward autonomous mobility.

The investment includes more than $7.5 billion to purchase robotaxi fleets and roughly $2.5 billion in equity stakes in autonomous vehicle companies.

Uber is partnering with multiple players across the ecosystem, including Baidu, Rivian, and Lucid Motors, as it builds a marketplace of robotaxi operators rather than developing its own self-driving technology.

The company plans to roll out robotaxi services in at least 28 cities by 2028, with investments tied to partners hitting deployment milestones.

Read more

-----------------------

74% of AI's Value Is Going to 20% of Companies

PwC released a global AI adoption study this week. The number that sticks is not the biggest one.

74% of AI’s economic value is being captured by just 20% of companies. The rest are stuck in pilots that never ship.

The gap is not about which model they use. The leading companies are 1.7x more likely to have cross-functional responsible AI governance in place.

That flips the most common boardroom argument. Governance is not slowing these companies down. It is the reason they can safely automate decisions at all.

Read more

Quick Hits

OpenAI acquires Hiro

A personal finance tool that managed over $1B in user assets. Product shutting down April 20. OpenAI is building toward high-stakes consumer decisions, not just productivity. Read more

Stanford 2026 AI Index

Private AI investment hit $150B globally in 2025. US leads in models. China is closing the gap in research and patents fast. Read more

FBI investigates attack linked to OpenAI CEO Sam Altman

A suspect has been charged after allegedly throwing a Molotov cocktail at Sam Altman’s home and later threatening OpenAI’s office, with the FBI calling the incident planned and targeted.

Ticketmaster brings event discovery into ChatGPT

Ticketmaster is integrating recommendations directly into ChatGPT to personalise how users find events.

Andon launches a marketplace for AI agents

Andon Labs has introduced a platform where users can buy, sell, and deploy task-specific AI agents, including tools for coding, research, and automation workflows.

Apple plans AI smart glasses strategy

Apple is developing smart glasses focused on real-time assistance and lightweight design, aiming to compete with Meta’s Ray-Ban lineup.

DeepMind shares new model capabilities

Google DeepMind highlighted updates to its models, including improved reasoning and multimodal performance in its latest release.

Tweet of the Week

AI desktop apps are becoming real workspaces

Anthropic and OpenAI are both updating their desktop apps, reportedly as early as next week.

The Anthropic version is said to bring multi-repository support and a cowork-style layout. These are starting to feel less like chat windows and more like actual development environments.

Before You Go

The PwC number will get screenshotted and shared a lot this week. The more interesting question it leaves behind is: what would it actually take for a company in the other 80% to close that gap?

It is probably not the model they are using.

See you next week! 👋