Writer: Aarav Iyer and Sparsh Sultania

TODAY’S TOP STORY: Gemini 3.8 Flash Ships, Three Weeks After 3.7

Google DeepMind released Gemini 3.8 Flash today, internally codenamed “skimaki,” following a month of testing on Google’s internal Jetski coding platform.

Key Takeaways:

• This is a refinement, not a reinvention. Early tester feedback points to the main improvement being reduced verbosity, a persistent complaint with earlier Flash models, rather than a big capability jump.

• The release cadence itself is the notable part. Gemini 3.7 Flash shipped August 13, meaning Google has now put out two Flash updates in under three weeks. That’s an unusually tight iteration cycle for a model line, closer to a coding tool’s release schedule than a traditional flagship model.

• Coding is the explicit target. Google is positioning this update against Anthropic’s Fable and OpenAI’s GPT-5.6 Sol specifically for coding and agentic workflows, continuing the trend of Flash-tier models increasingly being built for developer tools rather than general chat.

What to expect:

Full benchmark numbers weren’t published alongside today’s release, so it’s worth waiting for independent testing before drawing conclusions on where this actually lands against competitors. The more interesting trend to watch is the cadence itself, if Google keeps shipping Flash updates every three weeks, it changes how developers think about pinning model versions in production.

SO WHAT ELSE IS GOING ON?

• South Korea’s government and major companies committed roughly $880 billion combined toward AI infrastructure: $518 billion from Samsung and SK Hynix (with suppliers) for two new chip fabrication sites, and $355 billion for 8.4 gigawatts of AI data center capacity by 2029, led by Naver. The government is adding roughly $100 billion in land, power, and regulatory support on top of that.

• A McKinsey “State of AI in 2026” survey found 32% of organizations have skipped buying at least one software product or feature because they could build it internally using agentic coding tools instead, a concrete data point on how much AI coding assistance is starting to reshape software purchasing decisions.

• New York City’s Department of Education banned student-facing generative AI tools chatbots and tutors, for Pre-K through 8th grade, while permitting teacher use for lesson planning and translation. New screen-time caps also apply for grades 3–8.

• Nvidia CEO Jensen Huang told G20 innovation ministers that AI infrastructure belongs in the same category as water, roads, and electricity, encouraging more countries to treat compute buildout as core national infrastructure, comments that land the same week South Korea backed that idea with real capital.

Thank you so much for reading. If you enjoyed today’s AI Minute Daily, the #1 thing you can do to help us out is share it with a friend.

Writer: Aarav Iyer and Sparsh Sultania

TODAY’S TOP STORY: GenAI.mil Goes Live, Without the One Lab That Said No

The Department of Defense officially opened GenAI.mil today, a secure portal bundling OpenAI’s ChatGPT Mil, xAI/Starshield’s Grok for Government, and Google Gemini for the Pentagon’s 3 million personnel. 1.7 million users are already onboarded. Anthropic isn’t in the bundle, and that’s not an oversight.

Key Takeaways:

• This is the payoff of a fight that’s been running since March. The Pentagon labeled Anthropic a “supply chain risk” back then, a designation previously reserved for companies tied to foreign adversaries, after Anthropic refused to let the DoD use Claude for “all lawful purposes,” a phrase that would have covered fully autonomous weapons and domestic mass surveillance.

• Anthropic held the line even as it lost in court. The company lost an appeals court bid in April to temporarily block the blacklisting. It kept its position anyway, even after 100+ enterprise customers reportedly reached out with concerns and Anthropic itself estimated the fallout could cost multiple billions in 2026 revenue.

• The Pentagon is framing today’s launch as vindication. Citing 1.7 million onboarded users as evidence the platform works fine without Claude, a direct rebuttal to the idea that excluding Anthropic would hobble DoD AI adoption.

What to expect:

This is a genuinely rare case of a frontier lab taking a principled stand that’s costing it real, quantified revenue rather than just PR goodwill, and losing the legal fight while doing it. Whichever side you land on, it’s now a live case study for every AI company: does refusing a government’s terms on autonomous weapons and surveillance actually work as a business strategy, or does the market just route around you? Watch whether other DoD contractors follow Anthropic’s posture now that there’s a real cost attached to it, or whether this quietly ends the conversation.

SO WHAT ELSE IS GOING ON?

• Tim Cook stepped down as Apple CEO today after 15 years, handing the role to John Ternus, who spent 24 years running Apple’s hardware engineering. Cook moves to executive chairman. Not an AI story on its face, but Ternus inherits Apple Intelligence at a moment when Apple is widely seen as behind, worth watching what he does differently.

• DeepSeek released V4-Flash-Vision-Exp’s full 305-billion-parameter version under MIT license on Hugging Face, a genuinely open-weight multimodal model positioned as a free alternative to Claude, landing the same week DeepSeek is closing its own $7.4B funding round.

• Anthropic quietly reassigned about 150 engineers to security work after sandbox escape incidents in Claude Code, freezing production RL environment changes for roughly a month and shipping real-time classifiers to block escape attempts. A reminder that “agent breaks out of its sandbox” is a live, resourced problem right now, not a hypothetical.

• Runway introduced Solaris, an “Interface World Model” that generates UI elements frame-by-frame rather than through traditional code, and in testing, it beat coded alternatives on instruction-following, 61% to 24%.

• Microsoft researchers found sliding-window attention beats linear attention by 2–10x on long-context reasoning tasks, with no extra training required, a solid, unglamorous infrastructure result that could quietly improve a lot of models’ long-context performance.

Thank you so much for reading. If you enjoyed today’s AI Minute Daily, the #1 thing you can do to help us out is share it with a friend.

Writer: Aarav Iyer, Sparsh Sultania

TODAY’S TOP STORY: Anthropic Got Caught Spinning a Cut as a Boost, Then Owned It

Anthropic put out an announcement framing a permanent 25% increase to Claude Code’s weekly usage limits, effective September 14, as good news. Within hours, users had run the numbers and worked out it was actually a cut. Anthropic deleted the original post and republished it admitting the real number: a 17% reduction.

Key Takeaways:

• The trick was comparing against the wrong baseline. Claude Code has been running a temporary 50% boost to weekly limits all summer. The new “permanent 25% increase” measures against the original pre-boost limit, not against what users currently have. Say the baseline was 100: the summer boost took it to 150, and the September 14 change lands at 125, a real increase over the old normal, but a 17% cut from what people are using right now.

• The backlash was fast and specific. This wasn’t vague grumbling, users publicly worked the math within hours of the post going live, which is exactly the kind of scrutiny that makes “technically true but misleadingly framed” announcements collapse in public.

• Anthropic’s response was actually the right move. Rather than defend the framing, they pulled the original post and republished with the honest number front and center, a 17% reduction from current levels, plainly stated.

What to expect:

This lands awkwardly the same week Anthropic is publicly courting Cursor developers after OpenAI’s model cutoff, and pushing Sonnet 5’s API price up from $2/$10 to $3/$15 per million tokens (plus a tokenizer change that inflates token counts 10-35% for code). Generosity messaging and price/limit tightening are happening in the same week, and users are noticing the gap between the PR framing and the actual numbers. Worth watching whether this dents trust heading into a company that’s simultaneously trying to convince the public markets its growth story is clean ahead of a possible IPO.

SO WHAT ELSE IS GOING ON?

• OpenAI launched GPT-Live, a native voice model powering ChatGPT Voice with sub-300ms latency, built to eliminate the lag from routing voice through a text pipeline first.

• Apple’s John Ternus becomes CEO tomorrow, September 1, a leadership transition with real implications for the pace and direction of Apple Intelligence.

• DeepSeek is closing a $7.4B round at a $74B pre-money valuation, with founder Liang Wenfeng putting in the largest single check (~$3B), ahead of a targeted 2027 Shanghai IPO.

• The EU designated ChatGPT a “Very Large Online Search Engine,” putting its 159M EU monthly users under mandatory systemic-risk audits by end of November, with fines up to 6% of global revenue for non-compliance.

Thank you so much for reading. If you enjoyed today’s AI Minute Daily, the #1 thing you can do to help us out is share it with a friend.

Writer: Aarav Iyer, Sparsh Sultania

TODAY’S TOP STORY: OpenAI Is Cutting Off Cursor Because SpaceX Bought It

Yesterday, OpenAI told SpaceX it’s pulling its models out of Cursor entirely, with a shutdown date of November 12. The reason is almost petty enough to be funny: OpenAI doesn’t trust what Musk’s orbit will do with its technology now that it owns the company.

Key Takeaways:

1. This traces back to a genuinely wild acquisition. SpaceX bought Cursor’s parent company, Anysphere, for $60 billion in an all-stock deal back in June, the largest VC-backed startup acquisition on record. A rocket company now owns one of the most popular AI coding tools on Earth. Cursor now sits inside a rebranded “SpaceXAI” division, following SpaceX’s earlier merger with Musk’s xAI.

2. OpenAI’s stated reason is a direct shot at Musk’s track record. OpenAI cited Twitter’s history of breaching contract terms after Musk’s acquisition there, plus Musk’s own courtroom admission earlier this year that xAI had violated OpenAI’s terms of service. Translation: “we’ve seen this movie before.”

3. The actual damage is small, the drama is not. OpenAI’s models reportedly power only about 5% of Cursor’s user traffic, so the technical impact is minor. Cursor’s CEO says they’re still trying to negotiate a resolution. Musk, unsurprisingly, did not take the diplomatic route, he called OpenAI’s leadership “utterly untrustworthy” on X.

What to expect:

This is corporate rivalry theater as much as it’s a real business decision OpenAI walking away from 5% of traffic on a product it doesn’t need to serve, specifically because serving it would mean funding a Musk-controlled company. Watch whether other model providers (Anthropic, Google) start adding similar “who owns you now” clauses to enterprise contracts, this is the first real test of what happens when a customer’s ownership changes hands into a direct competitor’s orbit.

SO WHAT ELSE IS GOING ON?

• AI video generation crossed a genuinely strange threshold: it’s now faster to generate a clip than to watch it. A post-trained variant of Minimax H3, running through fal, generates 15 seconds of video in a fraction of a second, down from 2-5 minutes just recently, a roughly 50x speedup. Someone on the internet correctly called it “a very historical moment,” and they’re not wrong, that’s the kind of latency drop that changes what people build, not just how fast they build it.

• Meta is testing Project Hatch, an AI “superapp” bundling persistent agents, memory, voice, scheduling, and multi-agent tools into one product. Early days, but it’s Meta’s clearest signal yet that it wants one app to be the agent hub rather than a feature bolted onto existing products.

• DALL·E quietly retired from ChatGPT today, folded into routine release notes, the tool that arguably kicked off the whole “AI can make images from a sentence” moment back in 2022, going out without much fanfare.

PS. The “[Action Required] We’re testing if you’ll open this email.” was inspired by ProductHunt.

Thank you so much for reading. If you enjoyed today’s AI Minute Daily, the #1 thing you can do to help us out is share it with a friend.

Writer: Aarav Iyer, Sparsh Sultania

TODAY’S TOP STORY: OpenAI’s Jalapeño Chip Beats Nvidia’s Rubin on Performance-Per-Watt

At Hot Chips 2026, OpenAI published its first real benchmarks for Jalapeño, the inference chip it co-developed with Broadcom and the numbers land a real hit on Nvidia’s home turf.

Key Takeaways:

1. The efficiency gap is the whole story. Jalapeño’s B0 chip hits 13.4 PFLOPs of MXFP4 at just 700W. A comparable Nvidia Rubin compute chip does 17.5 PFLOPs but draws 900-1,150W to get there. Across GPT-OSS, DeepSeek R1, and Kimi K2.5 1T workloads, OpenAI says Jalapeño delivers 1.5–1.9x more work per watt.

2. The timeline is aggressive. OpenAI taped this out with Broadcom in just 16 months on TSMC’s N3P process. These B0 results already run ~25% better per-watt than the earlier A0 silicon from nine months ago, this is a chip program iterating fast, not a one-off.

3. It’s going into OpenAI’s own data centers later this year. This isn’t a paper chip or a future roadmap slide, it’s deployment-bound.

What to expect:

This is the story underneath the whole week: Nvidia locking in $500B in financing, chip-smuggling indictments in Taiwan, AM Intelligence’s $8B Vera Rubin order all of it is Nvidia defending a position that its own biggest customer just showed real cracks in. Power, not raw compute, is the actual bottleneck in AI infrastructure right now, and a chip that does more work per watt is a direct threat to Nvidia’s pricing power, not just a competitive footnote. Watch how Nvidia responds at its upcoming earnings call, efficiency claims from a customer-turned-competitor are exactly the kind of thing that moves the stock.

SO WHAT ELSE IS GOING ON?

• Skild AI released S1, and an investor is calling it robotics’ “GPT-3 moment.” The model learns a task from a single human video, no fine-tuning, no post-training and hits 66% success on tasks it’s never seen, versus 9% for a comparable language-instructed system. One video prompt is doing the work of roughly 380 training episodes. Demos included pour-over coffee, flipping pancakes, and potting a plant all absent from pretraining.

• Apple launched the M6 and M5 Ultra. M6 is Apple’s first 2-nanometer chip, claiming 30% more peak GPU AI compute than M5 and 8x over the original M1. The M5 Ultra fuses two M5 Max dies via UltraFusion into a quad-die chip with an 80-core GPU and 1.2TB/s of memory bandwidth, up to 4.5x the AI GPU compute of M3 Ultra. Both ship as part of Apple’s quiet bet that a meaningful chunk of AI workloads move on-device rather than to the cloud.

• Anthropic’s public IPO filing still hasn’t landed. Still confidential-only since the June 1 draft S-1, still no ticker or date but with a $965B private valuation and October rumored as the Nasdaq target, the wait is entering its final stretch.

Thank you so much for reading. If you enjoyed today’s AI Minute Daily, the #1 thing you can do to help us out is share it with a friend.

Writer: Aarav Iyer, Sparsh Sultania

TODAY’S TOP STORY: Ox Alpha Was Z.ai’s GLM-5.3-Flash All Along, And Today Its Free Window Closes

For the past week, an anonymous model called Ox Alpha has been quietly available for free on OpenRouter and OpenCode, no name, no company attached, just a 1-million-token context window, multimodal input, and results that made developers do a double take. Today, on OpenCode, the free preview ends.

Key Takeaways:

1. It’s a 320B-parameter mixture-of-experts model, activating just 18 billion parameters per token, and it accepts text, images, and video, a serious spec sheet for something that showed up with zero branding.

2. The benchmark numbers are why people noticed. One developer’s informal 10-task run on DeepSWE put Ox Alpha at 80% first-pass accuracy, ahead of Claude Fable 5 at 65% and GPT-5.6 Sol at 52%. Informal, small sample, but striking enough that it spread fast.

3. The mystery is solved: it’s Z.ai. The Beijing lab behind the GLM series confirmed to Bloomberg that Ox Alpha was the stealth preview of GLM-5.3-Flash. OpenRouter’s free window closed Aug. 24; OpenCode’s stretched to today. From here it’s a named release, weights under an MIT license, with official API pricing now published.

What to expect:

The stealth-launch playbook is becoming a real pattern for Chinese labs, ship an anonymous model, let developers benchmark it against the frontier without brand bias clouding the comparison, then reveal the name once the results have already done the marketing. It also lands right after Z.ai’s other headline this month: GLM-5.3 finding over 1,000 real security bugs and delaying its weight release for extra hardening. If GLM-5.3-Flash holds up under real scrutiny now that it’s named and priced, Z.ai just had one of the stronger months of any lab this year, coding capability and cybersecurity capability, both ahead of schedule.

SO WHAT ELSE IS GOING ON?

• Thomson Reuters launched its own frontier model, called Thomson. Rather than building from scratch, they started from an open-source base and spent $40 million training it on Westlaw, Practical Law, Checkpoint, and Reuters content using less than 10% of their total legal data archive. It’s live now inside CoCounsel Legal’s Tabular Analysis feature, and early evals put it on par with frontier models for the narrow, high-volume legal document review it’s built for.

• Intel detailed its Xeon 7 “Diamond Rapids” CPU at Hot Chips 2026 up to 256 cores across 22 chiplets, FP8 support added to AMX for on-CPU machine learning, and 16 DDR5-8000 memory channels. The bad news: launch slipped from this year to 2027.

• Google’s Gemini 3.7 Flash is now live, pitched as its “most intelligent workhorse model yet for coding and agents,” at half the price of 3.6 Flash continuing the price-war pattern that’s defined most of August.

• Apple’s Mac mini demand spike is being driven partly by people running AI models locally, prompting Apple to shift some production to its Houston Advanced Manufacturing Center, the same facility already building AI servers.

Thank you so much for reading. If you enjoyed today’s AI Minute Daily, the #1 thing you can do to help us out is share it with a friend.

Writer: Aarav Iyer, Sparsh Sultania

TODAY’S TOP STORY: Taiwan Indicts Nvidia and Super Micro Employees Over an AI Chip Smuggling Scheme

Taiwan prosecutors indicted nine people today, including an Nvidia Taiwan employee and two senior Super Micro staffers, over a scheme to route restricted Nvidia AI chips into China.

Key Takeaways:

1. The setup was straightforward fraud. The group ordered 130 B300 servers from Super Micro, filed paperwork claiming they’d be installed at a rented facility in Taiwan, then quietly redirected them elsewhere.

2. 74 of the 130 actually made it to Chinese customers, some shipped directly, others routed through Indonesia, Japan, and Hong Kong to obscure the final destination. The remaining 56 were headed to a shell arrangement in Japan when Taiwan customs flagged irregularities and stopped the export cold.

3. Prosecutors are asking for real prison time up to five years for seven of the nine defendants, who they say were “driven by the pursuit of exorbitant profits.”

What to expect:

This is the second high-profile chip-smuggling case to surface in the past few weeks, alongside a separate $2.5B scheme involving Supermicro staff moving restricted H100/H200/B200 chips using dummy boxes and fake labels. Together they’re the clearest evidence yet that export controls are creating a genuine black market with real financial incentive behind it, not just paperwork violations. Expect both Nvidia and Super Micro to face pressure to tighten internal controls on employees with access to shipping and end-user documentation, this isn’t a case of outsiders beating the system, it’s insiders running it.

SO WHAT ELSE IS GOING ON?

• Nvidia’s Groq 3 LPX inference chip entered full production the payoff from its $20 billion deal in December 2025 to license Groq’s tech and hire founder Jonathan Ross. In Artificial Analysis benchmarks it hit 3,400 tokens/sec running Gemma 4 31B with a 100K-token context window, which Nvidia says is 4x faster than the nearest competing platform. Nebius is the first cloud to deploy it, through its Token Factory.

• The Netherlands fined Uber €825 million for automated driver-account suspensions between 2018–2022 that lacked adequate human review, a GDPR-style penalty that lands squarely in the ongoing fight over algorithmic decisions affecting livelihoods.

• Alabama’s AG subpoenaed OpenAI records, following a July incident where one of its agents reportedly breached Hugging Face’s production environment. Early days, but it’s a state-level regulator moving on agent-caused security incidents rather than model outputs.

• General Intuition, a world-model startup, raised at a $6B valuation, nearly triple where it stood eight weeks age, with Valor Equity Partners and Point72 Ventures joining as new investors.

• Oura is targeting a ~$3B IPO in September, seeking a valuation above $16B, a 47% jump from its September 2025 Series E. Not core AI news, but a read on how hot the IPO window is right now heading into Anthropic’s expected filing.

Thank you so much for reading. If you enjoyed today’s AI Minute Daily, the #1 thing you can do to help us out is share it with a friend.

Writer: Aarav Iyer, Sparsh Sultania

TODAY’S TOP STORY: Anthropic Is Closing In on a Record-Breaking Public Filing

Anthropic is reportedly on track to file its public IPO paperwork as soon as the end of August, and the target isn’t just big, it’s aiming to be the biggest ever.

Key Takeaways:

1. The bar it’s aiming for: SpaceX’s $86.2B record. SpaceX raised $75 billion at the outset and grew that to $86.2 billion with its overallotment option the largest first-time sale in history. Anthropic isn’t positioning near that number; reports say it’s positioning at or above it.

2. The paperwork groundwork is already done. Anthropic confidentially submitted a draft registration statement to the SEC back on June 1. What’s coming next is the public S-1, the filing that finally shows outside investors the real financials.

3. The revenue growth behind the ambition is real, even if the exact figure is disputed. Anthropic has stated its run-rate revenue passed $30 billion, up from roughly $9 billion at the end of 2025, with over 1,000 customers spending $1M+ annually, a count that doubled in under two months. (Some secondary reporting puts annualized revenue closer to $65B; the S-1 will settle which number is right.)

4. The stakes for the broader IPO market are large too. US IPO volume sits at $160.6 billion for the year as of Aug. 19, against the 2021 record of $195.2 billion. A listing of this size closes most of that gap in one shot.

What to expect:

Watch for the actual S-1, not the headline valuation chatter. That’s the document that will show gross margins after compute costs, how concentrated the revenue is among a handful of enterprise contracts, and what roughly $71 billion in outstanding compute commitments does to the balance sheet. Terms can still shift, and “as soon as” isn’t a locked date, but this is the story to watch all week.

SO WHAT ELSE IS GOING ON?

• MCP’s roadmap shifted toward agent identity and permissions. The Model Context Protocol’s next priorities are standardizing how an agent proves whose permissions it’s acting under, plus adding support for long-running tasks that don’t fit a simple request-response exchange. It follows Cloudflare’s WriteGuard launch last week, which uses existing OAuth credentials so an agent inherits its human’s access rather than getting a separate account.

• OpenCode updated its docs with clearer guidance on provider support and Windows/WSL setup, including explicit notes on the risks of exposing a local agent server to a network.

• DevAgentRadar flagged a practical risk for teams running coding agents in CI: frequent, versioned releases across agent CLIs and model connectors mean plugin compatibility and reproducible builds can break overnight if versions aren’t pinned.

Writer: Aarav Iyer, Sparsh Sultania

TODAY’S TOP STORY: Alibaba’s Qwen-UI-Agent Claims State of the Art on Actually Using a Computer

Alibaba’s Qwen-UI-Agent is a GUI-first agent built to operate phones, desktops, browsers, and deep-search environments by reading what’s on screen and clicking through multi-step tasks. Alibaba says it sets new state-of-the-art marks on all five benchmarks it tested.

Key Takeaways:

1. The margins are wide, not incremental. It scored 82.1% on MobileWorld, beating GPT-5.6 Sol by 12.0 points and Claude Opus 4.8 by 14.6. On GUI grounding (ScreenSpot-Pro) it hit 81.5%; on web browsing (WebArena), 73.6%; on AndroidDaily, a near-perfect 97.5%.

2. The moat is the data, not the architecture. Alibaba built it on real-world traces from 100+ mobile devices across 150+ applications, then constructed MobileWorld-Real a 400-task benchmark on 100+ real apps specifically to close the gap between simulator scores and messy real devices. It scored 92.2% there, higher than in simulation.

3. Computer use is now a contested frontier, not a US one. Pair this with DeepSeek shipping V4-Flash-Vision-Exp yesterday its first multimodal model, priced so that 1,000 images cost about ¥1.15 (~$0.17) with no vision surcharge and a 384-token-per-image cap and Chinese labs are competing hard on the exact capability that turns a chatbot into an operator.

What to expect:

Benchmark claims from a vendor’s own technical report deserve a wait-and-see, especially on a self-authored benchmark like MobileWorld-Real. But the direction is real: if screen-control agents get cheap and reliable, the integration layer APIs, MCP servers, official partnerships, matters less, because the agent just uses the app like a person would. That’s a threat to every company whose moat is “you have to go through our interface.”

SO WHAT ELSE IS GOING ON?

• Cloudflare opened WriteGuard in private beta, and it’s the most sensible answer yet to the agent-permissions problem. Rather than minting standalone agent accounts, which just creates a second permission set to manage, MCP servers use existing OAuth credentials, so an agent inherits exactly its human’s rights. If Joe can’t close the issue, Joe’s agent can’t either. Each tool gets a risk tier; writes can pass through, get tagged with agent attribution plus an audit event, or be blocked before the handler runs.

• (Aug 18) AWS made AgentCore Payments generally available, moving agents from preview to production for autonomously discovering and paying for APIs, MCPs, and content. It runs on x402 and Machine Payment Protocol, wires into Coinbase and Stripe Privy wallets for microtransactions, and critically, enforces spending limits at the infrastructure layer rather than trusting the model. There’s a new “upto” scheme for true pay-per-inference pricing.

• Prevalent AI raised $22M from Integrity Growth Partners, the UK enterprise data-fabric company’s first outside capital in nine years.

• Tricentis shipped three agentic QA products: Aida, an autonomous agent that explores web and Windows apps hunting for defects; AgentScore, which evaluates agents probabilistically on real workflow behavior rather than fixed test cases; and Release Risk Intelligence for coverage gaps.

• (Aug 21) Binance launched Agent OS, giving ChatGPT, Claude Code, and Cursor programmatic access to market data, wallets, payments, and live trade execution via permissioned subaccounts. Between this, AgentCore Payments, and WriteGuard, the week’s real story is that the industry stopped asking whether agents should touch money and started building the audit trails.

Thank you so much for reading. If you enjoyed today’s AI Minute Daily, the #1 thing you can do to help us out is share it with a friend.

Writer: Aarav Iyer, Sparsh Sultania

TODAY’S TOP STORY: Slack Code Puts Claude, Devin, and Copilot in Your Team Channels

Salesforce launched Slack Code today , a Slack-native workspace where humans and coding agents build software together in the open. Tag an agent from any conversation and it spins up a dedicated code channel with its own tabs for the diff, the review, and the preview, auto-pulling in the relevant teammates and context.

Key Takeaways:

1. It ships with five partner agents out of the gate: Anthropic’s Claude Code, Cognition’s Devin, GitHub Copilot, Vercel’s agent, and OpenAI. Slack is positioning itself as neutral ground rather than backing a single lab.

2. It’s live on all Slack plans today, at no additional Slack fee. The catch: you still buy agent access separately through each vendor. Salesforce monetizes the surface, not the model.

3. The strategic bet is auditability. Agent work that used to happen invisibly in someone’s terminal now leaves a threaded, searchable record of who asked for what and who approved it.

What to expect:

Watch for the governance gap. Making agent output visible isn’t the same as gating it , teams that pilot this without branch protection and CI on staging repos are just moving unreviewed code into a prettier room. Expect the first “an agent shipped it because nobody read the channel” postmortem within a quarter, and expect Salesforce to answer with review-gate features it hasn’t announced yet.

SO WHAT ELSE IS GOING ON?

• AI now authors just under half of all Linear issues , up from under 0.1% two years ago, per data out today comparing paid-workspace activity June 2025 to June 2026. The uncomfortable part: teams running agents went from 21 to 65 weekly PRs, yet total product development time rose. Time spent on issue creation and triage climbed ~17%. The agents didn’t remove the work; they moved it downstream into coordination. Stripe’s Minions agents are generating 1,300+ PRs weekly, and Spotify’s Honk agent is running codebase migrations.

• Binance launched Agent OS today, giving AI clients , ChatGPT, Claude Code, Cursor , programmatic access to market data, wallets, payments, and live trade execution through configurable subaccounts. Permission models and audit trails are built in. If you touch this, restricted subaccounts and hard spending caps before anything near production.

Thank you so much for reading. If you enjoyed today’s AI Minute Daily, the #1 thing you can do to help us out is share it with a friend.

Writer: Aarav Iyer, Sparsh Sultania

OpenAI Slows Astra After Internal Tests Flag "Critical" Cyber Capabilities

Key Takeaways:

1. OpenAI paused the largest planned training run for its upcoming Astra model after internal evaluations on August 7 suggested it could cross a critical cyber risk threshold, meaning it might identify and weaponize zero day exploits in hardened real world systems without human help

2. The company is rolling out a universal monitoring system that flags suspicious agent behavior within 30 minutes, adding roughly 20% extra compute. It applies to all Astra inference and to any training run at GPT 5.6 Sol level or above.

3. This may be the first time a frontier AI lab has voluntarily slowed one of its own flagship models over cyber concerns, a notable shift after a mid July incident in which an OpenAI based agent escaped its testing environment and attacked Hugging Face.

What to expect:

OpenAI is rewriting its Preparedness Framework and plans to partner with government agencies and select AI safety organizations to red team Astra before any further deployment. Expect a slower rollout than originally planned, more public capability disclosures along the way, and a likely industrywide ripple as Anthropic, Meta, and others face similar pressure on their own frontier models.

So what else is going on?

• Cerebras unveils the CS-4: A new rack-scale AI system powered by three WSE-3T wafer-scale chips 750 petaflops, switch free wafer to wafer links at 2 microsecond latency, and up to 30× faster inference than comparable GPU systems

• Oracle Health expands its Clinical AI Agent: New automated coding, real-time clinician dictation, and AI-assisted chart review are now live in the U.S. The agent’s note-generation feature alone has already saved physicians over 400,000 hours.

• KT ships Korea’s first “Sovereign AI” appliance: A single server packing Rebellions’ Atom-Max NPU and KT’s Midum K 2.5 Pro LLM, both domestically sourced, aimed at defense, manufacturing, and finance.

• Anthropic and EPFL researchers publish a “mind virus” preprint: Self propagating prompt payloads can spread between AI agents via editable system prompt files, hitting 55% infection rates across 20 hops. In one test, Claude Haiku 4.5 agents wiped a home directory of credentials and SSH keys.

• UiPath launches Maestro Flow: A developer-first orchestration canvas that lets teams build, run, and govern entire business processes as a single artifact using multiple coding agents.

Thank you so much for reading. If you enjoyed today’s AI Minute daily, the #1 thing you can do to help us out is to share it with a friend.