The October model did not show up
A new OpenAI model was due inside ChatGPT in October, and in Codex beside it. According to The Wall Street Journal, that release is off.
Maxwell Zeff’s reporting, updated September 28, 2026, says the company is scrapping a model it calls GPT-6.1 Astra. Internal testing raised safety concerns. The Journal does not say what failed. It does say the debut those two products were waiting on is not happening on that calendar.
If you buy AI for business, that missed month belongs on the risk list. An automation plan that assumed October access, in the chat product or in the coding surface next to it, was planning against a date OpenAI, by the Journal’s account, did not keep.
The public record stops at safety concerns
The paper’s framing and its facts are not the same thing. The Journal calls the pull one of the clearest signs so far that agent misbehavior could stymie the industry’s rapid progression, and a rare case of a major AI developer ditching a new release because of safety concerns. Those are the paper’s characterizations. You will not find a lab notebook in them, or a description of what testers saw. Deception, tool misuse, data theft, or something duller: the reporting does not say.
A related clip on the same Journal page, about models going rogue online to steal information, is not a finding about GPT-6.1 Astra. Folding that clip into the cancellation is how a rumor becomes a briefing.
Some headlines use a punchier verb, including the claim that the model was benched for overstepping the mark. That wording is a headline. It is not the Journal’s. Until OpenAI publishes its own account, scrapping the release over safety concerns is the claim that has a named reporter and a date. Whether the model ships later is unanswered.
GPT-6.2, Astra, and Bel are not one product
Nothing in the reporting that documents this decision shows a GPT-6.2 release, a spec sheet, or a benchmark. Bel and Betelgeuse show up in the rumor stream. They are not a confirmed product. The Journal does not use those names for the model it says was pulled.
A different label is already on a competitor’s page. Anthropic’s write-up of Claude Opus 5.5 scores a model it calls GPT-6 Astra. Another label, GPT-5.6 Sol, appears in Anthropic’s coding comparisons and goes unexplained. Scoring something called GPT-6 Astra does not prove OpenAI has shipped it. It also does not prove that name is the GPT-6.1 Astra the Journal says was scrapped. The labels can be the same system, cousins, or a naming mess. Nothing public in these accounts settles it.
Refuse the vendor slide that collapses them into one GPT-6 you can buy on Monday.
A rival’s numbers are not a pilot
Anthropic’s Opus 5.5 page is the company’s marketing, plus the tests it chose to publish. Every percentage on it is Anthropic’s claim. No independent replication is available here. The page treats a model it calls GPT-6 Astra as a benchmark opponent. It does not say whether that model was released, whether it is the model the Journal says was pulled, or how Anthropic got the runs. If the opponent was unreleased or scrapped, the page does not show its work. The gaps below are vendor marketing against an opaque opponent. You cannot take them into a pilot as a cost-per-task finding.
The list price is narrower, and it is only a list price. Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5. A million output tokens is $20 at the new rate, $5 less than the previous Opus. That will not tell you what a finished task costs on your data.
One comparison is enough to see how the pitch is built. FrontierCode is a coding benchmark in that write-up. At what Anthropic calls default effort, it says Opus 5.5 scores 54.6 percent, just above the 53.3 percent it lists as GPT-6 Astra’s top, at about one-fifth the cost per task. A one-point gap and a cost claim, published by the seller, against a model these pages do not let an outsider identify. For AI automation, the useful lesson is the shape of the argument, not the score. If your pilot cannot measure cost per completed task on your own work, a borrowed percentage will not save the business case.
The math week is a different calendar
The other thread people keep stapling to “the next model” is a math result, not a chatbot launch. R&D World, citing OpenAI, reports that the company said in September its agents had found a solution to one forced-fluid version of the Navier-Stokes problem, one of the seven Millennium Prize Problems posed by the Clay Mathematics Institute in 2000. The Clay prize is $1 million. OpenAI said it would forgo it and published Lean files.
Plain version: start with fluid at rest, apply a smooth outside force, and produce a blow-up in finite time, speeds that become unbounded while total energy stays finite. R&D World says that matches a forced-motion version in Clay’s own statement. A machine-checked construction is not the same thing as mathematicians accepting that the prize problem is settled. The reporting keeps that distinction, and so should anyone repeating the claim.
The scale claim is OpenAI’s, via R&D World: roughly 10, 000 copies of an internal model, working as coordinated agents that could run code, read a cached snapshot of the web, and message one another. Further down the chain, R&D World cites a Unite.AI account of the announcement for the volume: launched September 1, a result about 88 hours later, 2.7 million messages, and about 130 billion output tokens. The formal files it points to are at github.com/openai/NavierStokesAndEuler.
On an October 29 episode of Stripe’s Cheeky Pint with John Collison, Altman put a number on that kind of exploration. What follows is R&D World’s rendering of the conversation, not a transcript checked against the audio:
I bet we could redo Navier-Stokes at 1/100th of the cost a year from when it was done.
R&D World reports that he called roughly $10 million the cost of working without a playbook, and that he said, “Look, we did let it get out of control for one week on the math. Intentionally so.” At the start, he said, they “understood absolutely nothing about the scaling laws of teams of agents.”
The same outlet reports a second forecast. It is not the same bet. Altman expects “10x improvements per year” in turning dollars or flops into useful work, from “smarter models and better approaches, ” not from already knowing the proof. A hundredfold cheaper rerun of a result the team already has, a year later, is not a standing promise of 10x gains every year. Do not average them into one efficiency number for a budget.
Three money figures are easy to mash together. One million dollars is the Clay prize. About $10 million is Altman’s figure, as R&D World reports it, for that exploratory week. “100s of millions of dollars of math” is how Collison put broader math spending. Altman agreed, and said the total “includes much more other discovery than what we have published, ” and that later problems cost far less. A headline that turns “a year from when it was done” into “by September 2027” is the outlet’s calendar, not a date in the quote.
A safety hold on a ChatGPT and Codex debut is one calendar. A research splash about teams of AI agents on a math problem is another. One is a product your teams might have been about to use. The other is a claimed result, a stated cost, and a bet. Neither is a GPT-6.2 benchmark.
What to plan against
The documented OpenAI event in this round is a model that did not meet customers. Next to it, a rival is already pricing work against a GPT-6 label. Sam Altman is on record, through R&D World’s account of a podcast, betting that a roughly $10 million agent week can be redone at a hundredth of the cost. Those are three different facts. They do not add up to a release note.
Keep AI automation tied to jobs you can score: a ticket closed, a brief that does not invent a number, a code change that survives the test suite. Keep a second model warm. The one on the roadmap can be pulled for reasons you will hear about late, and the public reason may be no more specific than safety concerns.
Codenames are free. Ten thousand agents for a week are not.
Questions worth asking before you replan
-
Did OpenAI just deliver its biggest leap in a model called GPT-6.2?
No. Nothing in the reporting that documents OpenAI’s decisions shows a GPT-6.2 release, a spec sheet, or a benchmark. Bel and Betelgeuse are names in the rumor stream, not a confirmed product.
-
What did the company actually withhold?
According to the Wall Street Journal, a model dubbed GPT-6.1 Astra, which had been headed for ChatGPT and Codex in October, after safety concerns in internal testing. The public record does not say what failed, or whether a later release is planned.
-
Is GPT-6 Astra the same model?
Unknown. Anthropic scores a model it calls GPT-6 Astra. The Journal writes about GPT-6.1 Astra being pulled. Using both names proves neither that they match nor that they don’t.
-
Did AI solve Navier-Stokes?
OpenAI has claimed a machine-checked result on one forced-fluid version and said it would skip the $1 million prize. That is not the same as the mathematical community accepting a full solution of the Clay problem. The $10 million figure is Altman’s cost, via R&D World, for a week of exploration. The hundredfold cheaper rerun is his bet, not a delivered invoice, and it is a different claim from the 10x-a-year expectation reported from the same conversation.
-
If a vendor slide says the product “beats GPT-6, ” what do you demand before it touches a forecast?
Four things, in writing. A model card, or an equivalent spec, for the model being sold and for the model it claims to beat. The access method: released product, private preview, or a run you cannot reproduce. Cost per completed task on your work, not a borrowed percentage. And a fallback if the roadmap model is pulled, because a safety hold can move a tool you were about to put in front of sales or engineering. A codename is not any of those four.