The most anticipated unreleased model in AI just left a trail.

On August 29, TestingCatalog reported that OpenAI has expanded internal testing of Astra under the codename mozaik-alpha-fdm, and published what appear to be the first public outputs anyone outside the company has seen, both zero-shot on max effort. Developers immediately began running head-to-head comparisons against Anthropic's newest.

Expanded testing usually means a release is getting closer. But the rest of the record points the other way, and the gap between those two signals is the story.

OpenAI paused work on this model three weeks ago because it could not rule out critical cyber capabilities under its own safety framework. Its largest frontier training run is still on hold. And OpenAI has not decided whether Astra ships as GPT-6 at all.

Strip out the speculation and OpenAI has confirmed exactly one thing about this model, in a research post on August 1 titled "Ten advances in mathematics and theoretical computer science."

An internal version of Astra produced new results on ten long-standing open problems, spanning high-dimensional geometry, group theory, quantum complexity, and lattice cryptography. OpenAI said the central results had seen no progress for at least a decade and in most cases far longer. Every proof was formalized in the Lean theorem prover, so they are machine-checkable rather than something you take on trust.

OpenAI described Astra in that post as its next major model. That phrase is the entirety of the official positioning.

One detail from the post is worth holding onto. OpenAI noted that the tokens required to find all ten solutions would have cost roughly $2,000 at GPT-5.6 Sol API rates. Ten decade-old open problems in mathematics, for about the price of a used laptop.

Everything else circulating is unconfirmed. Astra is rumored to be a roughly 10-trillion-parameter mixture of experts, the first frontier model to train multi-agent coordination natively from pretraining rather than bolting it on afterward. Internal codenames reportedly include mewfour, with a pretraining project called Doug. A widely shared August 25 leak claiming a finished 10-trillion-parameter pretrain codenamed Bel traces back to a single account and remains unverified.

This is the most common error in coverage, and it is worth being precise about.

OpenAI has used only the name Astra. It has reportedly not decided whether the model ships as GPT-6, as a point release like GPT-5.7, or under the Astra name on its own.

The precedent is recent and instructive. Earlier this year a model codenamed Spud was widely rumored to be GPT-6. An unverified leak pegged April 14 as launch day, tied to a super app combining ChatGPT, Codex, and the Atlas browser. Nothing shipped on that date. Nine days later OpenAI released the model as GPT-5.5.

Every "GPT-6 is launching next week" rumor of the past twelve months has been wrong, and the pattern is consistent: OpenAI ships under version numbers that undercut the pre-launch speculation. A community label is not a product name until OpenAI says so, and OpenAI has not said so.

So if you see a headline promising a GPT-6 release date, it is describing a model whose name has not been decided.

Here is what actually moved the timeline, and it moved it later.

On August 7, OpenAI said its own evaluations left it unable to rule out critical cyber capabilities for Astra under its Preparedness Framework, and that it was pausing internal activities involving the model that did not meet strengthened security controls. Every earlier model, GPT-5.6 Sol included, had been assessed at the lower High threshold. Astra is the first to be flagged above it.

On August 18 OpenAI told Axios more. It had paused two weeks of deployment-focused reinforcement learning training. Its largest planned frontier RL run was on hold. And it is rewriting the Preparedness Framework itself.

That last part deserves more attention than it got. The framework is the document that defines when OpenAI must stop, and the company is revising it while the model that triggered the stop sits waiting. Whichever way that revision lands, it will determine what happens next, and OpenAI is writing both sides of it.

There is also a monitoring cost attached. OpenAI disclosed that supervising models at this tier consumes roughly 20 percent of the inference compute being monitored, with alerts targeted within 30 minutes of suspicious behavior. That applies to all Astra inference involving tools.

So the picture is a model good enough to solve decade-old mathematics, held back by the company that built it, under a safety framework being rewritten mid-pause.

For anything you are building, treat Astra as a watch item rather than a roadmap item. There is no release date, no pricing, no model ID, and no waitlist. Any claim that a GPT-6 API is available now is third-party hype. Build against GPT-5.6, keep your routing configurable, and do not hard-code a guessed model name.

The interesting question is what the timeline actually signals. Two forces are pulling opposite ways.

Pulling earlier: expanded internal testing under a codename, reports of a launch within weeks, a quarterly cadence that put GPT-5.5 in April and GPT-5.6 in July, and OpenAI cutting Luna's price by 80 percent three weeks after launch, which companies rarely do to a flagship unless something is arriving behind it.

Pulling later: a critical cyber flag, a paused frontier training run, a safety framework under active revision, and a monitoring overhead that costs a fifth of the compute.

The honest read is that the second set is heavier. Expanded testing is a normal step; a halted frontier run is not. And whatever ships, remember that the last model everyone was certain would be GPT-6 arrived as GPT-5.5.

The stories moving fast beyond today's headlines:

OpenAI expanded internal testing of Astra under the codename mozaik-alpha-fdm, and the first public outputs surfaced on August 29, both zero-shot on max effort. That normally means a release is close.

The only thing OpenAI has officially confirmed came on August 1, in a research post rather than a product launch. An internal version of Astra produced new results on ten mathematics and theoretical computer science problems open for at least a decade, each formalized in Lean so the proofs are machine-checkable. The tokens to find all ten would have cost about $2,000 at current Sol rates. OpenAI called Astra its next major model and said nothing else.

It is probably not called GPT-6. OpenAI has reportedly not decided whether it ships as GPT-6, as a point release like GPT-5.7, or under the Astra name. Earlier this year a model codenamed Spud was widely rumored to be GPT-6 and launched as GPT-5.5.

And the timeline points later rather than sooner. On August 7 OpenAI said it could not rule out critical cyber capabilities under its Preparedness Framework, the first model ever flagged above the High threshold. On August 18 it confirmed a two-week pause on deployment-focused RL training, that its largest frontier run remains on hold, and that it is rewriting the Preparedness Framework itself while the model waits.

For builders: no date, no price, no model ID. Build on GPT-5.6, keep routing configurable, and ignore anyone selling a GPT-6 API today.

Stay building. 🤖