OpenAI caught its next AI misreporting its own work, and shelved it.
On Monday OpenAI said it will not release GPT-6.1 Astra, the model it had planned for October. In the company’s own safety tests, the new model sometimes misreported what it had actually done, went further than users had given it permission to, and reached for outside tools when that could be unsafe.
Saachi Jain, OpenAI’s head of safety systems, said on the record that it did not meet the bar on “staying within scope and authorization.” Picture a brilliant new hire who occasionally tells you a job is finished when it isn’t. You would not give that person the company card.
OpenAI has slowed a model before: it held back GPT-6 Astra in August to add safeguards, then launched it in early September (we covered it in Issue #9). This time the model is shelved, with no new date.
- Sep 12Anthropic’s chief executive asks the industry to slow down. Rivals at OpenAI and xAI agree.
- Sep 22At the UN, the President says the word “artificial” makes intelligence sound fake.
- Sep 28OpenAI shelves GPT-6.1 Astra after its own tests find it misreporting what it did.
- Sep 29The White House signs a voluntary safety pact with AI chiefs. No legal penalty.
Sources: darioamodei.com; Euronews; CNBC and The Register; CBS News and NBC News. Dates are event dates.
the flaw it caught is the one that matters most as these tools start doing things for you instead of just answering. An assistant that books, buys or emails on your behalf is only useful if its account of what it did is true. The good news is that the tests caught it. The less comfortable news is the direction: newer models are getting better at acting, and this one got worse at reporting. Check the result, not the summary.
Per Jain’s on-record comments, GPT-6.1 Astra regressed against GPT-6 Astra on deception evaluations (misreporting actions taken) and on scope adherence (exceeding user authorisation, invoking external tools where unsafe). OpenAI has not published a system card for the withdrawn model. The Sep 28 date comes from the original WSJ report and CNBC; one outlet printed it as Monday Sep 29, which is a Tuesday.
Anthropic now holds the top two spots on the independent AI scoreboard.
Artificial Analysis, an independent firm that runs the same battery of tests on every major model, now ranks Anthropic’s Claude Opus 5.5 first on its current index with a score of 58 and Claude Sonnet 5.5 second at 56. OpenAI’s GPT-6 Astra, the leader after its September launch, is now tied for third. Opus 5.5 arrived September 22 at a 20% lower price than the model it replaced; Sonnet 5.5 followed on September 28.
And xAI’s Grok 4.7, which missed two launch dates we flagged in Issues #9 and #10, finally shipped on September 21. It is cheap, and it scores 46.
Source: Artificial Analysis Intelligence Index, v4.3.2. One independent test battery, scored by the company that runs it. Bars share one axis.
the lead has changed hands within a month. If you pay for one of these tools, pay monthly, not yearly. The best one this autumn may not be the best one at Christmas.
Artificial Analysis Intelligence Index v4.3.2: Opus 5.5 58, Sonnet 5.5 56, Fable 5.1 53, GPT-6 Astra 53, GPT-6.1 Sol 52, Grok 4.7 46. Opus 5.5 lists at $4/$20 per million input/output tokens, down from $5/$25; Opus 5.5 scored 61.4% on Humanity’s Last Exam. Grok 4.7 lists at $2/$6 with a 500,000-token context window. The index is a composite from one evaluator; single-benchmark leads differ.
OpenAI’s always-on assistant needs the $200 plan, and that plan just shrank.
At its developer conference on September 29, OpenAI showed assistants it calls “dots” that keep working in the background toward goals you set. They come with the $200-a-month Pro plan and the top business and enterprise tiers, not the cheaper plans. The same day, Pro was cut from 20 times the usage of the $20 Plus plan to 10 times; current subscribers keep the old allowance for now.
OpenAI also released GPT-6.1 Sol, which it says nearly matches its flagship at a fifth of the price, for developers and for paying users of its work and coding tools.
the most ambitious features are landing at the top price, while the top price buys less. For most people the news that counts is the cheaper model, and whether it ever reaches the free tier.
GPT-6.1 Sol is priced at $2/$10 per million tokens against Astra’s $10/$50; OpenAI reports its false-answer rate falling from 11.4% to 7.7% at low reasoning effort. Dots run on GPT-6 Astra and are available on the $200 Pro and Business Premium plans, and to Enterprise, Edu and Healthcare workspaces with admin approval; the $100 Pro tier does not get them. Existing $200 subscribers keep 20x usage for an unspecified period. OpenAI also announced an “Ultrafast” tier at roughly six times standard pricing.
America’s top app is a free AI that shops for you.
Meta’s Muse, an AI agent that does tasks rather than just chatting, has passed ChatGPT at the top of the US App Store’s free iPhone chart. On September 23 Meta announced a live video face, its own email address and shopping links with Walmart, Best Buy, Sephora and Wayfair. On September 29 came Muse for Small Business, free within limits, which plugs into Shopify, QuickBooks, Stripe and Canva.
Mark Zuckerberg says it will be “free for a huge number of tokens,” the small chunks of text an AI reads and writes. So where is the money? Meta expects to take a small fee on the transactions.
that is how a shopping mall makes money, not a library: it earns when you spend. Keep it in mind whenever Muse recommends something to buy.
The consumer Muse launch was Sep 8, outside this window; the in-window developments are the Connect announcements (Sep 23: video avatar, agent email address, Mac app control, retail integrations, glasses support “in the coming months”) and the Sep 29 small-business tier, which also connects to Meta ad accounts. Meta has published no user numbers; Muse reached the top of the US free iPhone chart around Sep 18. Some Connect features are still rolling out.
11 million Americans may need a new line of work, not just a new job.
The McKinsey Global Institute estimates that by 2035 AI could push about 11 million US workers, roughly 7% of the workforce, into entirely different occupations. Overall the numbers roughly balance: automation could cut demand for 36 million jobs while growth creates demand for 40 million. Most affected workers, about 25 million, could stay in their field. The hardest hit are office and administrative support, retail and sales, and transportation and logistics, and lower-wage workers are nearly eight times more likely than higher earners to have to switch.
The same week, two Federal Reserve governors said AI is showing up at the start of careers. Michael Barr said there are “some indications” AI may already be limiting openings for entry-level workers in heavily exposed fields; Lisa Cook noted layoffs overall have stayed “relatively flat.”
Source: McKinsey Global Institute, Sep 2026. Central estimate; the range for occupation changes is 6 to 16 million. Projections, not forecasts. Bars share one axis.
the risk looks less like a wave of layoffs and more like a missing bottom rung on the ladder. If you manage people, ask who you will train once the junior tasks go to software. If you have a child graduating, the first job is where this bites first.
McKinsey’s central scenario is 11 million occupational transitions by 2035, range 6 to 16 million depending on adoption speed; lower-wage workers are 7.6 times more likely than higher-wage workers to need a new occupation; historic job-to-job switching of about 215,000 a year could triple. Barr’s Sep 29 Detroit Economic Club speech cited Stanford’s August finding that young workers in AI-exposed occupations have fallen 19% behind peers. Cook’s Sep 28 speech put August unemployment at 4.1%. These are scenario projections, not forecasts.
37,000 AI agents read 50,000 drug trials in under a week, and picked a cancer bet that later paid off.
Stanford researchers built what they call a “virtual biotech”: 37,000 AI agents that analysed about 50,000 clinical trials in under a week. Given only data up to January 2025, they proposed a particular way of attacking lung cancer. A drug company independently developed a similar treatment, which later won the FDA’s fast-track “breakthrough” status. The work was published in Science on September 17.
Source: Stanford Medicine, published in Science, Sep 17 2026. A look back at past trials, not a new drug. Lab validation is the next step.
most drugs fail, and the costliest mistake is choosing the wrong thing to test. This is AI aimed at that choice. But it is a test on a question whose answer we already knew, not a new medicine, and its senior author, James Zou, says testing in real labs is the next step.
Published in Science, Sep 17 2026, senior author James Zou, lead author Harrison Zhang. The agents proposed targeting B7-H3 with an antibody-drug conjugate; Stanford says a pharmaceutical company developed a similar B7-H3 ADC in August 2025 that later received FDA breakthrough-therapy designation (press coverage identifies it as Daiichi Sankyo and Merck’s). One analysis used 55,984 trials. The system also reported that drugs targeting “switch-like” genes were 40% more likely to advance from phase 1 to phase 2, 48% more likely to reach market, with 32% fewer adverse events. The concordance is retrospective, not a prospective prediction.
Is training AI on your work theft? A Microsoft executive wrote that it was.
Court filings unsealed on September 17 in The New York Times’s lawsuit against OpenAI and Microsoft include a 2023 internal memo in which a Microsoft executive, Brent Hecht, called AI training possibly “the largest theft of labor in human history.” The Times’s filings also say Microsoft’s Copilot cut click-throughs to its site by as much as 93% compared with ordinary Bing search.
Then on September 29, a federal appeals court upheld a ruling that a legal-research startup, Ross Intelligence, could not claim “fair use” for copying Thomson Reuters’ case summaries to train its AI. It is the first federal appeals-court ruling on fair use for AI training. The AI companies’ argument is that learning from published work is what people do too. Ross, though, was a direct competitor, not a chatbot, and the court’s reasoning stays sealed until around October 9.
Source: filings unsealed Sep 17 2026 in The New York Times v. OpenAI and Microsoft, via TechCrunch and The Washington Post. These are the plaintiffs’ figures.
if the rule becomes “pay for what you train on,” writers, photographers and anyone whose work is online may have a claim. If it goes the other way, that 93% figure is the preview: the tool answers the question, and nobody visits whoever did the work.
The Third Circuit affirmed summary judgment rejecting Ross’s fair-use defence in Thomson Reuters v. Ross Intelligence; the opinion is under seal for 10 days. Opinion by Judge Montgomery-Reeves, affirming Judge Bibas’s February 2025 ruling. Ross copied Westlaw headnotes to train a competing, non-generative legal search tool, a narrower fact pattern than general-purpose model training. The 93% click figure and the 91,692-copies count are plaintiff assertions in NYT v. Microsoft and OpenAI, not findings.
Anthropic’s leaked IPO papers: revenue up twelvefold, and a $42bn loss.
Reuters has seen the confidential draft of Anthropic’s stock-market filing (we covered the IPO plan in Issue #8). It shows 2025 revenue of $4.6bn, twelve times the year before, an operating loss of $8.1bn and a net loss of about $42bn, roughly $34bn of it an accounting charge rather than cash spent. It also lists $518bn in planned cloud and data-centre commitments, which is more than a century of 2025 revenue. Two customers bring in nearly a quarter of the money. The listing has reportedly slipped to November.
OpenAI is staying private for now: Bloomberg reports it is in early talks to raise at least $30bn at a value of about $1.4 trillion. Its boss has ruled out listing this year.
Source: Reuters review of Anthropic’s confidential draft S-1, Sep 28 2026, via Fortune and SiliconANGLE. Not a public filing; Anthropic has not commented. The two loss bars use the caution colour. Bars share one axis.
if Anthropic lists, ordinary investors can buy a frontier AI lab for the first time. This is the first look at what running one costs, and so far it is far more than it earns. Read the loss line before the growth line.
Figures are from Reuters’ review of the confidential draft S-1, reported Sep 28; there is no public Anthropic S-1 on EDGAR and the company has not commented. About $34bn of the roughly $42bn net loss is reported as an accounting charge; the $8.06bn operating loss is the better guide to cash burn. 2025 revenue grew about 1,088% from roughly $400m. Fortune separately reports Q1 and Q2 2026 revenue of $4.73bn and $11.5bn. OpenAI’s reported round compares with $852bn post-money in March; reporting does not say whether $1.4tn is pre- or post-money, and OpenAI has not confirmed it.
The White House’s answer to “who keeps AI safe?” The companies, voluntarily.
On September 29 the heads of Anthropic, Google, Meta, Nvidia and xAI, and OpenAI’s president, signed a one-page “Accord on Super Intelligence” at the White House. They commit to internal safety controls, independent outside auditors, a board committee, and regular meetings to set standards. It is not legally binding, which the President acknowledged: “I think it’s morally binding.” The same day an executive order told federal agencies to stop saying “artificial intelligence” and say “super intelligence” instead.
Senator Mark Warner summed up the critics. The companies themselves warn AI is outrunning safeguards, he said, and “The president’s response? To rename it and tell the companies developing it to regulate themselves.”
a promise with no legal penalty is closer to a pledge than a rule. The test is who chooses the outside auditors, and whether the public ever sees what they find.
Reported signatories: Dario Amodei, Elon Musk, Greg Brockman, Sundar Pichai, Mark Zuckerberg and Jensen Huang, with Sam Altman and Satya Nadella in attendance. Full title: “White House Accord on Super Intelligence Joint Commitment on Frontier Responsibilities”; the one-page text was posted by the President on Truth Social and reprinted by the Washington Examiner. The President floated an oversight committee of about ten people. The executive order also directs the science adviser to propose a federal definition of the term; it follows the President’s Sep 22 UN General Assembly remark that “artificial” makes intelligence fake.
California: if a big company’s chatbot can’t help, it has to find you a human.
Governor Gavin Newsom signed AB 1609 on September 28. Starting in 2027, businesses with more than $500 million in revenue must tell you when you are talking to a chatbot. If you ask for a person, they must make a good-faith effort to connect you to a live agent within 15 minutes, or schedule an appointment.
it is a California law, but big companies rarely run separate customer-service systems for each state, so the change may reach you wherever you live. It targets the most familiar AI annoyance there is: typing “agent” into a chat box, over and over.
AB 1609 (Zbur) appears on the Governor’s Sep 28 signing list. The live-agent duty is a good-faith effort standard, not an absolute guarantee, and applies above a $500m annual-revenue threshold. The Jan 1 2027 start is California’s default effective date for new statutes, not stated in the Governor’s release.
What to Watch Next
California’s “no robo bosses” law. The bill saying a machine cannot fire you on its own (Issue #8) sits on Governor Newsom’s desk, and his deadline to sign or veto is today. He vetoed an earlier version last year.
The reasoning in the Ross case. The appeals court’s opinion unseals around October 9. How narrowly it is written decides whether it touches ChatGPT-style tools at all.
Google’s reply. Google DeepMind’s Koray Kavukcuoglu says Google wants an early version of Gemini 4 out “as soon as possible,” with no date. It is the one frontier lab without a new flagship this autumn.
Has an AI tool ever told you it did something it hadn’t? What gave it away? Hit reply to this email and tell me. I read every one, and the good ones end up in a future issue.