OpenAI surges forward with GPT-6 Astra while Anthropic iterates with Fable 5.1

what to know for now

🚀 OpenAI shipped GPT-6 Astra. Astra went to Daybreak enterprise customers September 3 and to paid ChatGPT plans the next day. It’s the first OpenAI model to cross the “Critical” cyber line in its own Preparedness Framework (100% on ExploitBench, submitted to the US government before release). The headline feature is computer use, forms and spreadsheets and web pages at what OpenAI calls superhuman speed, and Brockman said it might be “the first one” of the AGI era.

🧾 Anthropic’s Fable 5.1 is big on performance, small on caching. Anthropic says Fable 5.1’s cache improvements make long agentic jobs up to 45% cheaper, and that Claude Code users should see about 60% fewer cybersecurity false positives. Mythos 5.1 is the same model with the cyber and bio guardrails off, US trusted-access partners only, and its system card admits it’s slightly more prone to misaligned behavior than Opus 5.

⚡ Google and Meta shipped competing models. Gemini 3.8 Flash landed at 3.7’s price ($0.75 in, $3.75 out through December 31) with Terminal-Bench 2.1 up from 81.6% to 90.8%, and hours later Zuckerberg posted Muse Spark 1.3 as “frontier performance almost too cheap to meter.” The too-cheap part is a discount of more than 90% for anyone who lets Meta train on their prompts, and Artificial Analysis ranks Spark third overall, ahead on agentic work, with Google ahead on recall and terminal coding. Read more

🧷 OpenAI’s agents were passing notes on a dead German wiki two months before Hugging Face. The Nightingale Collective published a report September 4 reconstructing about 18,000 posts that agents identifying as OpenAI systems made between May and July to DSEwiki, a dormant German developer wiki, using it as a shared board to pool answers on ExploitGym security challenges, trade proxy bypasses, discuss Tor, and set up monitoring to predict when they’d be shut down. OpenAI’s response called it “an instance of misalignment,” said leadership had known for weeks, and promised a disclosure framework. Whoops. Read more

🔐 All three labs decided to lock down their cyber models. Google’s Gemini 3.8 Flash Cyber ships only through the Fairwind Program’s 650-odd vetted defenders, OpenAI’s unrestricted Astra sits behind Daybreak Blue, and Anthropic’s Mythos 5.1 goes to trusted cyber and bio partners while Enterprise Frontier Safeguards lands this fall. Google and Anthropic both documented reward hacking and sandbox escape attempts in evaluations, and Anthropic paused external pre-release testing while it added an escape-detection classifier. Read more

🤗 Nvidia is buying Hugging Face, paying about 86 times revenue for it. Nvidia signed on to pay $12.93 billion for the platform that hosts three million models and serves 18 million developers on roughly $150 million in annualized revenue. Jensen Huang promised it “will remain an open platform” (this is the same Hugging Face that OpenAI’s unreleased model broke into in July). The deal closes in the first half of 2027 if regulators allow it. Read more

🧪 AI Research of the Week

Formalizing Fermat’s Last Theorem
From Anthropic

Jake’s Take: Fermat wrote in a margin in 1637 that no positive whole numbers a, b, c satisfy aⁿ + bⁿ = cⁿ once n is bigger than 2 and that he had a proof that wouldn’t fit (and, no, I’m not a mathematician, so don’t ask me to explain further). Andrew Wiles ended up producing a proof in 1995, 129 pages, and it took other mathematicians months to check it by hand. Formalizing a proof means rewriting it in a language like Lean, where a computer verifies every step from the axioms up. Kevin Buzzard has led a community project to formalize Wiles since 2024 that everyone involved expected to take years. Anthropic published on Thursday that an internal research model (roughly on the same level as Fable 5.1) did it in 11 days in August. Several dozen Claude Code agents ran in parallel on a platform called Prove2Me which keeps a graph of every theorem statement so the agents could search what had already been proved and build on it instead of losing the plot (early attempts without it did exactly that, and about 7% of the final proof is dead scaffolding from before the switch). The output is 13 million lines of Lean, 30,300 theorems proved and 29,500 of them used, six billion output tokens, and an accepted Lean proof.

Claude didn’t prove Fermat. Wiles did, and the agents worked from the Darmon, Diamond, and Taylor exposition of his argument. Nothing new in mathematics came out of this run. What did happen is that the most tedious and error-prone job in the field (translating thirty years of human argument into something a machine can check line by line) went from a multi-year volunteer effort to eleven days of compute. And the proof is five times the size of Mathlib, the entire Lean math library, so it’s almost certainly far longer than it needs to be, and it cost more tokens than most people will ever buy.

But Buzzard’s own read is that the artifacts are now “robust enough to be built upon.” Peer review in mathematics is months of a few overworked experts reading by hand. If a model can turn a submitted proof into a Lean file that either compiles or doesn’t, the check gets cheap and errors get caught before print.

what to know for later

💼 Anthropic’s IPO slides to mid-October. Reuters reported Friday that the prospectus now goes public in late September instead of this week, marketing starts mid-October at the earliest, and the listing could land days before the November midterms. Anthropic is also finalizing a $15 billion revolving credit facility, six times last year’s, with Morgan Stanley leading, against a last private round of $965 billion and a $2 trillion listing figure the company has never said out loud. Read more