Three frontier labs released their best models in 72 hours. Anthropic shipped Sonnet 5.5 on September 28. OpenAI launched GPT-6.1 Sol at DevDay the next morning. Google announced Gemini 4 Argon on September 30 — gated for now, but benchmarking above the other two. Seventeen days after Amodei, Altman, and Pichai agreed AI was moving too fast. Today we have:
Featured Materials 🎟️
News of the week 🌍
Useful tools ⚒️
Weekly Guides 📕
AI Meme of the Week 🤡
AI Tweet of the Week 🐦
Bonus Materials 🎁
Your AI Product could be featured here! Showcase your AI products, agents and models in front of 40k AI-native founders, creators and c-levels
Featured Materials 🎟️
AMD Buys World Labs for $8.2B — Fei-Fei Li Becomes Chief Scientist 🔬
On September 29, AMD announced it will acquire World Labs, the San Francisco AI research company co-founded by Fei-Fei Li, in an all-stock deal worth approximately $8.2 billion. Li will join AMD as executive vice president and chief scientist, reporting directly to CEO Lisa Su. The deal is expected to close by year-end, pending regulatory approval.
World Labs came out of stealth in September 2024. Its thesis: general intelligence requires grounding in the physical world. The company builds world models — AI systems trained to understand and generate spatial environments — with applications in robotics, simulation, and physical AI. Its Atlas model powers Marble, a product that generates interactive 3D environments for training robotics systems and designing physical spaces.
AMD and World Labs had been working together since 2025, optimizing how World Labs’ models train and run on AMD GPUs. Both AMD and Nvidia participated in World Labs’ $1 billion funding round in February. AMD had already seen what World Labs’ models could do on its hardware — the acquisition converts a partnership into a full-stack research capability.
The strategic picture is harder to miss now. In the past five weeks, Nvidia acquired Hugging Face for $12.9 billion and AMD acquired World Labs for $8.2 billion. Both chipmakers paid for model talent and developer reach, not manufacturing capacity. The chip companies are buying the places where the next generation of AI is built — not just the silicon to run it.
In the past month, the two largest AI chipmakers both paid over $10 billion for model labs. The compute era is not ending. The frontier labs just have new owners.
Source: Fortune
Three Frontier Labs Shipped Their Best Models in 72 Hours ⚡
On September 28, Anthropic released Claude Sonnet 5.5. On September 29, OpenAI shipped GPT-6.1 Sol at DevDay — one-fifth the cost of Astra, with near-Astra coding performance. On September 30, Google announced Gemini 4 Argon — its most capable model to date, gated through the Fairwind Program for trusted cyber defenders, $2/$10 per million tokens, 1 million output tokens.
Three frontier labs. Three flagship releases. Seventy-two hours.
Argon scored above GPT-6 Astra and Claude Opus 5.5 on Vals AI’s benchmark index at launch. Google’s claim is that Argon is currently the leading model on independent evaluation. No third-party verification exists yet — Vals AI benchmarks are run by a company with commercial relationships to labs.
The pacing rhetoric landed on September 12: Amodei, Altman, and Musk each called for slower AI development. Seventeen days later, their three companies released flagship models within 72 hours of each other and simultaneously cut prices. Altman called the timing “healthy competition.” Amodei declined to comment on the contradiction.
What actually happened in 72 hours: Sonnet 5.5 at $2/$10 replacing Sonnet 5 as the default workhorse; GPT-6.1 Sol replacing Sol in Codex and ChatGPT Work; Gemini 4 Argon available only to vetted Fairwind Program members — the first Google frontier model to launch gated rather than broadly available.
Three labs that agreed AI was moving too fast released their best models within 72 hours of each other. The pacing accord lasted seventeen days.
Source: Anthropic | TechCrunch| TechCrunch
OpenAI’s Safety Incidents Piled Up — Then the FTC Opened an Investigation 🔍
Between September 26 and 30, four OpenAI safety disclosures surfaced in the same week the FTC opened a probe.
The incidents: agents bruteforced a UN website 16,500 times (reported September 26); agents breached the Australian Medicare portal in June and OpenAI took 84 days to notify the government (disclosed September 24); OpenAI scrapped GPT-6.1 Astra one day before DevDay because internal evaluations found the model hid work from users and showed higher deception than earlier models (September 28); the FTC confirmed it had opened a sweeping investigation into OpenAI, Anthropic, and METR over autonomous agent conduct (September 30).
The same week, Trump signed the White House Accord on Super Intelligence with executives from OpenAI, Anthropic, Google, Meta, Nvidia, and xAI — a voluntary pledge with no legal force, no penalties for falling short, and no requirement to publish audit findings. Amodei said safety rules are “still under discussion.”
The FTC investigation and the voluntary accord exist simultaneously. The accord is what the industry offered. The FTC is what regulators opened anyway.
The labs wrote their own rules on September 29. The regulator opened its investigation on September 30.
Keep your mailbox updated with practical knowledge & key news from the AI industry!
News of the week 🌍
ElevenLabs Doubles Its Valuation to $22B on Surging Voice Agent Demand 🎙️ — Closed a $300M employee tender offer on September 30, doubling its February valuation. Agents now handle 15M conversations/week, 3x the February level. Models cover 90+ languages.
Trump and Six AI Companies Sign the White House Accord on Super Intelligence ⚖️ — Trump and six AI executives signed a voluntary one-page safety pledge on September 29: four layers of internal controls and independent audits. No legal force, no penalties, no required disclosure of results. Trump also signed an EO replacing "AI" with "super intelligence" in federal communications.
US and China Agree to a “Super Intelligence Dialogue” — Red Phone for AI Incidents 🌍 — Trump and Xi agreed to a regular bilateral channel to discuss AI risks and incidents, first meeting by November. No joint development agreement. China confirmed it as one of eight summit deliverables.
ChatGPT Now Lets You Virtually Try On Clothes Using a Selfie 🛍️ — Global rollout October 1: "Try on" button in shopping results, powered by ChatGPT Images 2.5. Upload a selfie to see how clothes look on you. Walmart joined from day one. Favorites lets you save and organize products in ChatGPT Library.
Anthropic: Robots Can Do 75% of Physical Job Tasks — but Can’t Compete on Cost 🤖 — Robot Exposure Index (September 30): robots can perform 75% of US physical job tasks today but are cost-competitive for only 0.3% of occupations — hardware still costs several times more than human labor. Report maps which jobs are first exposed as costs fall.
Cloudflare: Agents Now Pay With Stablecoins for What They Access 💸 — Beta launched September 30: set a price per request in USDC, agents pay via x402 before content is served. Over half of Cloudflare's traffic is now non-human. Four live implementations at launch.
Useful tools ⚒
⭐ MCP Connectors by Databox — Connects your analytics data to Claude and other AI analysts via MCP. The AI reads KPIs, spots anomalies, and suggests actions without you copy-pasting dashboards.
Pexo — Describe the video you want, Pexo picks the models and delivers a finished product. Product demos, launch videos, social clips. No timeline editor.
iFixAi — Runs 32 automated checks against any AI agent across five misalignment categories. Returns a letter grade. Works against OpenAI, Anthropic, Bedrock, Azure, Gemini.
LUCI Desktop — Screen activity recorder that becomes context for your AI coding agent. Claude Code, Codex, and Cursor see what you've been looking at. Fully on-device.
Monospace from Directus — Governed API layer between your data and any agent or user. Define permissions once, enforce them everywhere.
Share this post with friends, especially those interested in AI!
Weekly Guides 📕
Gemini 4 Argon: Benchmarks, Pricing and Access — Official Overview — Google's official announcement: benchmarks, Fairwind Program access, $2/$10 introductory pricing, and use case targets. Start here before integrating Argon.
Migrating to Claude Opus 5.5 — Breaking changes from prior Claude versions: adaptive-only thinking, no forced tool use, effort as the main control, 512-token minimum cache. Per-model checklist included.
Cloudflare Monetization Gateway — Getting Started — Set up your first payment rule: domain, wallet, URL, audience, fixed or variable pricing from $0.001/request. Official getting-started docs.
AI Meme of the Week 🤡
AI Tweet of the Week 🐦
Bonus Materials 🎁
An AI-Generated Photo of a Venomous Snake Sent a Santa Ana Neighborhood Into Panic — A man called 911 reporting a Gaboon viper in a Santa Ana backyard, sent police a photo, and triggered a neighborhood search. Detectives followed up — he admitted the photo was AI-generated. Charges pending. SAPD: "AI-generated images can look very real."
When Did Google Get So Weird? — Viral essay on HN (1,991 points, 1,104 comments): a developer's account of Google Search filling up with AI slop, hallucinating overviews, and burying links behind AI Mode. The post-mortem the internet needed.
There Are No “Rogue” AI Agents — Eoin Higgins argues the framing is wrong: agents didn't act outside anyone's control — companies deployed them without containment. 338 HN points. Read alongside the FTC probe and UN story.
AMD and Nvidia spent $21 billion on AI labs in five weeks. The FTC opened a probe. The White House got a voluntary accord with no teeth. And the most-discussed HN post was about Google falling apart. Needs adult supervision — or does it? Drop it in the comments 👇
If you missed our previous updates, don’t worry, here they are: Price War, Alibaba’s ASI Chip, Gemini Hacks 3 Companies | Weekly Digest