Two labs appear to have their next generation finished. Neither will tell you when you can use it.

Leaks circulating this week claim Anthropic has completed internal training on its next foundation model, launching as either Mythos 6 or Mythos 6.1, with one report claiming it outperforms Mythos 5 by 12.5 percent on Cobbench v2. OpenAI confirmed its next model family is called Astra two weeks ago, not with a launch but by publishing ten solutions to mathematics problems that had been open for at least a decade.

No dates. No pricing. No availability. From either of them.

The reason is the interesting part, and it is not that the models are not ready. Let us go through what is actually known, what is rumor, and one claim making the rounds that does not survive checking.

Mythos is not a chatbot. It is the foundation model underneath Anthropic's systems, which is why it matters more than a product release. Improvements there eventually surface across Claude, Claude Code, and everything built on top.

The claims are that internal training is complete, that final branding has not been settled between Mythos 6 and 6.1, and that an unreleased model scores 12.5 percent above Mythos 5 on Cobbench v2.

Treat all of that as rumor. Anthropic has confirmed none of it, there is no model card, no API string, and no independent benchmark. The one detail that reads as credible is the training completion claim, because it fits a pattern Anthropic has already demonstrated. Mythos 5 itself leaked in March through a CMS misconfiguration, months before anyone could use it, and Anthropic confirmed the model existed while saying nothing about a date.

The more substantive part of the leak is the explanation for the delay. It claims deployment of newer Anthropic models has been slowed by restrictions surrounding Fable 5 and Mythos 5. Your readers know that history: those are the two models the US government export-controlled in June, taking them offline worldwide for eighteen days.

OpenAI's situation is the same shape with better documentation.

Astra is real and named by OpenAI itself, revealed in the third paragraph of a research post rather than a product launch. The claim is that an internal version produced new solutions to ten problems in mathematics and theoretical computer science open for at least a decade, spanning high-dimensional geometry, group theory, quantum complexity, and coding theory, including a construction proving non-sofic groups exist that answers a question posed in 1999.

What separates it from a press release is that OpenAI formalized the logic in the Lean theorem prover and published it, making the proofs machine-checkable rather than something you take on trust.

Astra is described as multiple agents coordinating over hours or days on a single problem. Altman demonstrated it to policymakers in Washington. It is expected to be among the first models submitted for federal pre-release review. And there is still no release date, no pricing, no API, and no statement on whether it becomes GPT-6.

Since it is circulating widely, worth addressing directly: the Opus 6 rumor does not hold up.

A tracker that follows Anthropic's model lineage checked it and found no corroborating evidence in any leaked API string, any Cursor model-picker sighting, or any reporting from established trackers. Their assessment is that it is a six-month-old identifier being recirculated rather than a fresh signal, matching the pattern of an earlier false leak.

Two structural reasons support that. Anthropic has moved Opus by decimal increments, 4.5 through 4.8 and then 5, rather than whole-number jumps. And its newest capability tier went into the Mythos and Fable class above Opus, not into a numbered Opus 6. If something more capable than Opus 5 is coming, the naming history says it arrives as Opus 5.x or as a Mythos or Fable release, not as Opus 6.

Repeat the rumor if you like, but that is where the evidence sits.

Put both labs side by side and the bottleneck comes into focus. It is not capability.

Anthropic's most capable models were export-controlled by the US government in June and went dark worldwide for eighteen days. Even after controls lifted, Fable 5 moved in and out of subscription availability repeatedly through the summer. Astra is expected to go through federal pre-release review under the framework the administration has been negotiating with the labs since GPT-5.6 was gated at launch. And more than 1,100 employees across OpenAI, Anthropic, Google, and Meta signed a letter last month asking Washington to build the tools to deliberately slow frontier development.

That is a very different industry than the one that shipped GPT-4 the week it was ready.

There is a commercial layer too. Anthropic is targeting a listing in late September or early October, and OpenAI's has reportedly slipped toward 2027. Neither company wants a capability disclosure or a safety incident landing in the middle of a roadshow, and both have now published incidents where their own models breached containment during evaluation.

So both have their next generation finished, and both are moving carefully for reasons that have nothing to do with whether the model works.

The stories moving fast beyond today's headlines:

Mythos 6. Leaks claim Anthropic has completed internal training on its next foundation model, launching as either Mythos 6 or 6.1, with one report claiming a 12.5 percent gain over Mythos 5 on Cobbench v2. None of it is confirmed, there is no model card or API string, and the only element that reads as credible is the training completion claim, since Mythos 5 also leaked months before availability. The leak attributes the delay to restrictions surrounding Fable 5 and Mythos 5.

Astra. OpenAI confirmed the name two weeks ago inside a research post claiming ten decade-old mathematics problems solved, with the logic formalized in Lean so the proofs are machine-checkable. It runs multiple agents coordinating over hours or days, Altman demonstrated it in Washington, and it is expected among the first models submitted for federal pre-release review. Still no date, price, or API.

Opus 6 is not real. A model tracker found no corroborating evidence in any API string, Cursor sighting, or established tracker, and concluded it is a recirculated six-month-old identifier. Anthropic moves Opus by decimals rather than whole numbers, and its top tier went into the Mythos and Fable class rather than a numbered Opus 6.

The through-line. Both labs appear to have their next generation finished and neither can ship it, which means the bottleneck on frontier AI has moved from training compute to regulatory clearance. That advantages labs with government relationships and compliance infrastructure, and it explains why Chinese labs are shipping every two weeks while American frontier releases stall. Build on what has an API today.

Stay building. 🤖