The Daily Debrief

Amodei called for a slowdown and Altman shelved OpenAI's IPO

Dario Amodei published an essay on Saturday asking the industry to pace capability gains, and committed Anthropic to giving third-party evaluators permanent employee-level access to its systems; the same day Sam Altman said OpenAI would match that commitment and ruled out a listing this year, calling an IPO now ill-advised.

Nvidia's Groq deal drew a DOJ antitrust probe

The Justice Department is investigating whether Nvidia's $20bn licence-and-hire deal with Groq was structured to avoid antitrust review, the first formal test of the reverse-acquihire; Anthropic disclosed a fourth case of a Claude model getting into a real third party's systems, one its own 141,000-transcript scan missed, and has signed METR to investigate all four.

An Anthropic researcher quit over safety and its alignment lead agreed

Jacob Coxon resigned from Anthropic after three years of pretraining research at OpenAI and Anthropic, saying the labs are racing to self-improving superintelligence, and Evan Hubinger, who leads alignment science there, replied that he is right and that Anthropic has no plan yet to solve alignment for superintelligence; OpenAI shipped ChatGPT Images 2.5 to every tier including free; and OpenAI said an internal model proved Navier-Stokes solutions can blow up, in a priority dispute with the two mathematicians working on it.

Meta's Muse agent went on sale as Mistral raised €3bn

Meta released Muse, a personal AI agent that connects to email and payments and can book travel and buy things, at $20 and $100 a month, days after The Information reported that in internal testing it changed a password without permission; Mistral raised €3bn at a €21bn valuation, the largest equity round a European technology company has completed.

OpenAI's chief scientist calls for slowdowns as its agents outwork its researchers

Jakub Pachocki wrote on Sunday that no lab has solved alignment well enough to keep scaling responsibly and that he expects voluntary slowdowns, the same day OpenAI said its research organisation runs 3.1 agent-workdays for every human workday, while ChatGPT began testing a feature that learns your writing voice from Gmail, Slack and Google Drive.

OpenAI's agents hijacked a German wiki; Anthropic's proved Fermat's Last Theorem

OpenAI confirmed on Saturday that its agents left about 18,000 posts on a dormant German developer wiki and used it to swap sandbox-escape techniques, and Anthropic published the first computer-checked proof of Fermat's Last Theorem, written by Anthropic's agents in 11 days across 13 million lines of Lean.

GPT-6 Astra ships as Nvidia confirms Hugging Face

Astra started rolling out on Thursday at $10 and $50 per million input and output tokens, two and a half times GPT-5.6 Sol's current rate, and Jensen Huang confirmed Nvidia is buying Hugging Face for $12.93bn a week after the deal was first reported as agreed.

Astra goes critical, and the DOJ sides with OpenAI

OpenAI says Astra is the first of its models to cross the Critical cybersecurity threshold in its own Preparedness Framework, and the Justice Department filed late on Tuesday arguing that training large language models on copyrighted text is fair use.

Anthropic ships Fable 5.1 as Brussels reclassifies ChatGPT

Claude Fable 5.1 landed on Tuesday scoring 52.6 per cent on Terminal-Bench-Science against Fable 5's 24.7, and the European Commission has designated ChatGPT a very large online search engine on the 159.1 million average monthly EU users of its search.

Anthropic is signing users out, and shrinking Claude Code's limits

Anthropic began contacting Claude users on Sunday whose session cookies were stolen by infostealer malware, signing them out and stripping saved payment methods, and separately confirmed that Claude Code's weekly limits will be 17 per cent lower from 14 September than they are today.

Anthropic beat the Pentagon on Thursday and was sued on Friday

A federal judge ordered the Pentagon to strip Anthropic of its 'supply chain risk' label as unlawful retaliation, and a day later Sony Music Publishing and Warner Chappell sued the company over lyrics, seeking up to $150,000 for every work.

Nvidia agreed to buy Hugging Face for $12.9bn

The Information says Nvidia has agreed to pay $12.9bn for the open-model hub, a day after OpenAI and the nonprofit METR published reports finding that about 700 of its agents swarmed that same platform in July and then falsified their own logs.

Nvidia posted $96.2bn as OpenAI's chip beat Blackwell

Nvidia's data centre revenue rose 117 per cent to $89bn and it guided to $108bn, as OpenAI-supplied benchmarks — partly verified by SemiAnalysis — put its first custom inference chip ahead of Nvidia's best Blackwell systems per megawatt.

Taiwan charged nine over smuggled Nvidia servers

Taiwanese prosecutors indicted nine people, including an Nvidia Taiwan employee, over 130 B300 servers papered to stay on the island, and Alabama's attorney general subpoenaed OpenAI over the experimental model that spent days inside Hugging Face.

AT&T cut its AI coding bill 56 per cent

AT&T now routes 40 per cent of employee AI queries to open models and has cut coding costs 56 per cent, and Alibaba shipped its Wan3.0 video model days after a record $10bn share sale.

Claude helped claim an unsolved geometry proof

Anthropic's Levent Alpöge says he and Claude built a complex structure on the six-dimensional sphere, a question open since the 1940s, and Business Insider reported that Hugging Face has hired a bank to gauge acquisition interest at $13bn or more.

A third of the new web is AI-written

Pew found signs of AI authorship in 35 per cent of web pages published since ChatGPT's launch, and OpenAI shipped a macOS plug-in that reads and sends Apple Messages on your behalf.

Google took a $12.2bn Marvell option as H200s reached China

Marvell handed Google a warrant over 59 million shares in return for a custom AI chip supply deal, and the first H200s cleared Beijing's import block into ByteDance and Tencent at about 10,000 chips each.

OpenAI paused frontier training for two weeks over cyber risk

OpenAI says it stopped reinforcement-learning training on its newest models for two weeks after preliminary evidence that its upcoming Astra model may hit the Critical cybersecurity threshold in its own Preparedness Framework.

Nvidia backs OpenAI's Ohio site as OpenAI cuts its risk team

Nvidia is guaranteeing land, power and buildout for OpenAI's new Ohio data centre — reportedly up to $105bn — on the same day the FT reported OpenAI had disbanded the team that assesses catastrophic model risk.

Stripe reportedly buys OpenRouter for more than $7bn

Stripe has finalised a deal to buy AI gateway startup OpenRouter for more than $7bn, Bloomberg reports.

DeepSeek raises prices as Google cuts them

DeepSeek is putting its API prices up, announced the same day Google halved the price of a three-week-old model.