Nobody has put a source on next Thursday, and nobody put one on the Thursday before it. Two questions circulate about Astra: whether it ships on a named day, and what Anthropic does when it finally does. The first is answerable, and the answer is that the day doesn't exist.

A specific weekday attached to an unreleased frontier model is either a leak or a game of telephone, and this is telephone. Chase the word back through the coverage and it arrives at a video published on Wednesday 19 August, titled "GPT-6 Astra Releasing Thursday". Its own description is markedly softer, saying Astra "could be arriving as soon as Thursday" on the strength of X posts by OpenAI employees. The title hardened what the description hedged, the whole mechanism inside a single artefact, and the Thursday in both was the next day. Before that, on the 6th, a leaker named a release candidate and said Astra was days away, and by the time I wrote that up on the 15th nothing had shipped. A fresh date arrives about as fast as the last one expires.

Set against that, the company's own record runs to three posts. The 1 August research post named Astra "our next major model" and credited an internal version with ten long-open problems in mathematics and theoretical computer science. On 7 August it said its evaluations meant it "cannot rule out critical cyber capabilities" under its Preparedness Framework. On the 18th it said a significant number of workloads remain paused, which reads to me as containment rather than restraint. Not one of the three carries a date, a price or a model ID, and a White House official told Axios that OpenAI had volunteered its plans to delay.

The capability underneath the rumours isn't in doubt, which is most of why they regenerate. The Lean repository for those ten proofs reports a sorry count of zero, meaning no step in any of them is left unproven, and it's a compiler saying so rather than a press office.

TestingCatalog reported on 29 August that internal testing has expanded around a checkpoint called mozaik-alpha-fdm, with the first outputs circulating publicly, while noting OpenAI has confirmed neither the codename nor a date. Real evidence of preparation. The betting markets priced on 18 August gave 18% for a launch by end of August; with the month nearly out and no model, that leg resolved the way the pattern predicts. The same book had 60% by 15 September, a date that keeps sliding rightward.

Anthropic's answer was written in February. The same Axios piece notes it had committed to pausing training if capabilities outran its ability to control them, then removed that in a Responsible Scaling Policy update, on the reasoning that one developer pausing while others "moved forward training and deploying AI systems without strong mitigations" could leave the world "less safe". That commitment covered training rather than launches, so it doesn't map straight onto a deployment hold. The reasoning generalises, though: Anthropic has written down that it won't be the lab that stops while a competitor keeps moving.

So a slowdown is the one response the framework rules out. On timing I have less. One unconfirmed report from the Chinese outlet 36kr has Anthropic's Fable 5.1 finished and held to ship within hours of GPT-6, and by the standard I've applied to every Thursday here, it doesn't clear the bar either. It doesn't have to. The doctrine gets there without it.

Sources: