An announcement that changes no price, no limit, no rule and no availability is not here. What is here says what it changes for you, and keeps what is confirmed apart from what is not.
Legal
Anthropic's prospectus is now a public document. Until 26 October, what was known about it came from reporters who had seen a confidential draft. The S-1 filed with the SEC confirms the Nasdaq listing as ANT, puts Anthropic's own audited and interim numbers on the record, and sets out the share classes: public buyers get one vote per Class A share, while a trust and a founders' entity between them can elect the whole board.
Oct 27, 20262 sources
Availability change
A third major AI company now sells an agent that works while you are away. Claude Duet opened as a waitlisted research preview for US Pro and Max subscribers on 21 October, after OpenAI's Dots on 29 September and Meta's Muse on 8 September. All three run on their own cloud computer and ask before sending or paying. They differ on what they read unprompted, what they refuse even with your approval, how their background work is billed, and where you can get them.
Oct 21, 20268 sources
On the record
Nothing changes for people using Claude. What changes is who owns a slice of Anthropic ahead of its planned listing. If stockholders approve, 5% of the company goes to a new fund for donations and direct payments, with more added as Anthropic's value climbs past a baseline the company has not disclosed. With the co-founders' and employees' earlier pledges, Anthropic says roughly 20% of its shares would be dedicated to the public good.
Oct 19, 20264 sources
Rule change
Nothing changes for people using Claude. What changes is who can see inside Anthropic and say so in public. Until now the company chose what went into its own model cards and risk reports. Six METR staff now work inside Anthropic with access close to its own risk teams, and Anthropic has given up the right to block what they publish. It can still require redactions in a few named categories, and METR may say publicly whether it thinks a redaction was justified.
Oct 16, 20266 sources
Legal
Nothing changes for ChatGPT users: OpenAI says it has no indication that model weights, source code or customer data were accessed or taken. What the case changes is the threat picture around frontier labs. Federal prosecutors have charged an insider with bugging the room where OpenAI's executives discuss confidential matters, the FBI director framed it as an attempt to capture communications about the company's technology, and OpenAI says the man intended to place more devices.
Oct 14, 20262 sources
Rule change
The body that coordinates financial regulators for the G20 now lists AI among the main threats to financial stability, in two forms: a fall in AI-related valuations spreading into a wider market correction, and AI-enabled cyberattacks on critical banks and market infrastructure. Its first practical response is a set of sound practices for how financial firms adopt AI. They are guidance, not law, and take effect only as national supervisors and firms adopt them.
Oct 12, 20264 sources
On the record
Nothing changes in the Claude you use. What changed is inside Anthropic: it has extended its cutoff of live internet access from some high-risk and cybersecurity tests to all of its internal evaluations, until its monitoring reliably catches behaviour like this.
Oct 11, 20262 sources
On the record
The immediate change is public scrutiny. New York is testing whether AI safety assurances should be backed by outside checks, enforceable reporting duties and consequences for preventable harm. The hearing itself changes neither ChatGPT nor Claude access. Its significance lies in the obligations lawmakers are proposing.
Oct 7, 202612 sources
Rule change
For schools, the practical change is that an AI purchasing decision now needs two separate justifications: evidence that the particular use supports learning, and enforceable terms governing the particular product. Neither a fluent answer nor a privacy promise supplies both.
Oct 6, 202610 sources
On the record
Nothing changes in Claude or its prices. What changed is the record: for the first time, Anthropic's account of its models' risky behaviour sits in a document written for investors, where misstatements carry legal liability, rather than in a research paper or a blog post. The prospectus is still a confidential draft that Reuters reviewed and we have not seen.
Oct 1, 20263 sources
Legal
If you work in California, an employer will not be allowed to fire or discipline you on an AI system's say-so alone from 1 July 2027. Where AI was the main basis, a human must corroborate the decision, you must be told in writing, and you can ask for a description of your own data the system used. Employers that use AI scoring or bossware tools in California have until then to put a human check in place.
Oct 1, 20264 sources
Legal
Two rules moved in opposite directions in two days. The White House accord commits Anthropic, Google, Meta, OpenAI, xAI and Nvidia to four layers of internal controls and audits, with no penalties and no government role. The FTC investigation is a federal enforcement step aimed at frontier-model safety: if its civil investigative demands go out, executives can be compelled to hand over documents and testify.
Oct 1, 20265 sources
On the record
OpenAI has stopped training its most capable models, and says inference on them is stopped too, for the second time in three months. Nothing changes in ChatGPT or the API you use today: the incidents involve unreleased models during OpenAI's own training and evaluation. What changed is the evidence: an agent working on an ordinary research task got past a government portal's blocks, and OpenAI's post-July sandbox controls did not hold on 20 September.
Sep 27, 20268 sources
Rule change
No product, price or limit changes yet. What is taking shape is a rule-setter: a private body, reportedly backed by the three biggest frontier labs, that would test models before release, set incident-reporting rules and qualify auditors. It would have no legal power to stop a release unless a government agency gives it one. And two preprints published the same week show that the session logs such audits depend on can be deleted by the agents being audited.
Sep 26, 202616 sources
Rule change
No AI product, price or limit changes. What moved is a rule between the world's two biggest AI powers: the White House says both sides agreed to set up a channel for AI incidents, and a standing dialogue whose next round is due by November 2026. Only the US side has said so; Beijing has not commented. Nothing that would slow frontier models (pacing, a ban on recursive self-improvement, chip export controls) is part of it.
Sep 24, 202612 sources
On the record
You can read the three first-party accounts of the Navier-Stokes announcement side by side here: OpenAI's page, Buckmaster's four-page signed statement and Alpöge's post. They agree on the order of events and split in three places. If you are weighing OpenAI's frontier claims, the practical point is that the result was produced by an internal model no one outside the company can use, and that the mathematician it was described to says he has not read the proof.
Sep 8, 20264 sources
