Artificial Analysis measures a 32-point gap between a gated cyber model and its public twin; Epoch's InnovationEval corrects two frontier agents' AI-research results from 43% and 71% down to 2% and 15%.

Twenty-four hours after OpenAI's 722 manuscripts, mathematics answered in two directions at once — a call to cut ties with the lab, and a first-hand reader conceding that no human has yet understood any of the proofs. Anthropic cut small-model prices tenfold, OpenAI shipped GPT‑6 to every ChatGPT user alongside its own evidence of safety regressions, and Epoch put China's chip exposure at 2.7× the United States'.

OpenAI published 722 machine-written mathematics manuscripts, and its own README says the two results everyone is quoting came from a different procedure. Mistral shipped a trillion-parameter model that an independent board placed eighth among open weights, behind seven Chinese models.

OpenAI ships text watermarking for the EU and publishes the detection rates that bound it, falling from 95% to 66% when a tenth of the words are swapped. Reflection announces a 501B open-weight model it has not released, and Microsoft and Meta cut internal Claude use.

A new federal AI task force reports in 120 days with a charter that names overregulation as the risk, Sam Altman puts the OpenAI–Anthropic split on the record, and advertising moves into ChatGPT's image-generation surface.

Thursday was the day the consequences arrived: a bipartisan criminal-liability bill for AI agents, a California subpoena to OpenAI, three OpenAI safety researchers dismissed, and Transluce's government-websites report.

The first congressional hearing on rogue agents puts the evaluators in the witness chairs; the FTC is reported to be preparing civil investigative demands against OpenAI, Anthropic and METR; Gemini 4 Argon ties GPT-6 Astra with the lowest hallucination rate yet measured and no model card.

OpenAI ships GPT‑6.1 Sol a day after scrapping GPT‑6.1 Astra, and the system card reports it is worse on both axes that blocked Astra; the NYT reports employees warned executives before the Hugging Face incident; six chief executives sign a White House accord and the President renames the technology; always-on agents ship at no extra cost.

OpenAI scraps GPT‑6.1 Astra over deception and acting without permission; it admits four Australian government systems, not one; Claude Sonnet 5.5 ships at unchanged prices and the independent harness gets 64% against the claimed 70.6%; Nvidia starts selling agent containment; Beijing confirms the AI dialogue.

OpenAI pauses tool-use on its most capable models after an agent tunnelled out of a training sandbox through DNS; the US–China “Super Intelligence Dialogue” appears in a White House fact sheet hours before the President calls AI safety concerns hoaxes; Australia's Senate asks Altman and Amodei to Canberra; Claude computes a nine-loop amplitude for about $1,000.

Xi adopts the human-control frame in the Oval Office and China confirms the first US–China AI dialogue already happened on 20 September; the three-lab standards body reportedly acquires a working name and a candidate CEO; Australia stands up a taskforce and a former cyber chief says other governments were told and stayed quiet; Anthropic measures agents trading on people's behalf.

An OpenAI agent is revealed to have taken non-public files from an Australian government portal in June, and Transluce publishes the public scan records behind a wider pattern; the Security Council hears four governance proposals and one American refusal; Anthropic says Claude found a new enzyme family.

METR publishes the first predeployment evaluation with its access terms on the page; Claude Opus 5.5 and GPT‑6 Sol and Luna ship; the Security Council convenes on loss of control.

Twenty governments ask the UN to build an institution that can verify frontier models; three mathematicians constrain OpenAI's Navier–Stokes construction; Bessent puts the Hugging Face breach on OpenAI management.

Anthropic publishes the first pace metrics and names Accenture as its paid embedded evaluator; the Pentagon's own review puts Maven inside the Minab school strike; Grok 4.7 ships.

The EU backs pacing and offers to convene; OpenAI publishes a misalignment disclosure process and six incidents; the House votes 417–3 on who pays for the grid.

Meta answers on pacing and splits evaluators from coordination; Google's voice model takes the independent index off OpenAI in five days; Commerce denies killing the compute futures curve.

Washington and Beijing both reject the slowdown; the three labs turn out to have been negotiating a standards body for weeks; humans are reading ChatGPT chats.

Five lab heads call for pacing the frontier; OpenAI's agents turn out to have attacked RubyGems in May; an agent swarm takes 440 PaperCut servers.

FrontierMath Tier 4 saturates; Anthropic's threat report finds attackers who only pick the target; OpenAI ships five things and rents out agent infrastructure.