OpenAI Astra is a model you cannot use, announced in a way almost nobody expected. On August 1, 2026, OpenAI published a paper containing ten new results in mathematics and theoretical computer science, every one of them shipped with a machine-checkable proof. The name of the model behind the work, Astra, appeared partway down the announcement rather than in the headline. There was no launch event, no pricing page and no release date.
That combination makes Astra harder to assess than a normal model launch, and it has produced a lot of confused reporting. This article covers what OpenAI actually published, what the ten results establish and what they do not, why the widely repeated cost figure is misleading, how Astra differs from Google’s similarly named Project Astra, and what the new federal review framework does and does not require. Where a claim comes from OpenAI rather than an independent source, it is labelled as such.
The Key Takeaways
- Astra is not released. There is no release date, no pricing, no API access and no ChatGPT rollout. OpenAI has not said whether it ships as GPT-6 or inside the GPT-5 line.
- Ten results, ten machine-checked proofs. Each result ships with a Lean 4 certificate in a public repository, built with mathlib on Lean 4.32.0.
- The headline result is an explicit construction of a non-sofic group, settling whether every countable group admits finite permutation approximations.
- The $2,000 cost figure is incomplete. It covers the successful runs. OpenAI has not said how much compute was spent on attempts that failed.
- No government approval is required. The June 2026 executive order sets up a voluntary review with no power to block or license a model release.
What OpenAI Announced on August 1
OpenAI published a 249-page manuscript titled Ten Advances in Mathematics and Theoretical Computer Science, along with a set of reasoning walkthroughs and a public repository of formal proofs. The paper’s abstract describes the work as “a collection of results obtained by an internal OpenAI model.” The name Astra does not appear in that abstract at all. It appears in the accompanying blog post, where OpenAI identifies Astra as its next major model.
This detail matters more than it sounds. Astra has no model card, no API documentation and no product page. What exists is a research artifact and a name attached to it. Anyone telling you Astra has been launched, benchmarked or priced is working from something other than OpenAI’s published material.
The supporting repository is public and verifiable. It was created on August 1, 2026, carries an Apache 2.0 licence, and contains ten Lean files, one per result. Gizmodo characterised the announcement as being smuggled into a blog post about mathematics, and that reading is fair. A model family reveal was placed inside a research write-up rather than given its own launch.
What OpenAI Astra Actually Proved
The ten results span high-dimensional geometry, coding theory, group theory, operator algebras, arithmetic circuit complexity, quantum complexity, lattice cryptography and extremal combinatorics. Three of them resolve numbered problems from the Erdős problem collection. Below is the full list in plain language.
| # | Result | What it establishes |
|---|---|---|
| 1 | Sphere packing | Pins down exactly how strong the Cohn-Elkies method can get, improving the general packing bound in high dimensions |
| 2 | Binary and spherical codes | Improves classical upper bounds by exponential factors at every parameter setting |
| 3 | Non-sofic groups | Constructs an explicit non-sofic group, settling whether every countable group admits finite permutation approximations |
| 4 | Connes’s rigidity conjecture | Disproves it, by building infinitely many non-isomorphic groups sharing one von Neumann algebra |
| 5 | Arithmetic circuit complexity | New lower bounds for computing the permanent with circuits and formulas |
| 6 | Quantum parallel repetition | Extends a classical repetition principle to every finite two-player entangled game |
| 7 | Closest vector problem | Proves polynomial-factor hardness, with consequences for lattice cryptography |
| 8 | Ehrhart’s volume conjecture | Proves the sharp volume bound in every dimension |
| 9 | Multicolor Ramsey numbers | A superexponential lower bound, resolving Erdős problem 183 |
| 10 | Compactness and degeneracy | Disproves two conjectures in extremal graph theory, resolving Erdős problems 146 and 180 |
The headline is number three. Soficity asks whether every countable group admits approximation by finite permutations, and no explicit counterexample existed. Astra produced one, using property-(T) expanders and the binary Leavitt algebra. Number four is comparable in weight, disproving a long-standing conjecture of Alain Connes in operator algebras.
Humans were involved, and OpenAI has been clear about where. Company researchers turned the model’s raw output into publishable manuscripts, while OpenAI states that the mathematical arguments themselves came from Astra. That division matters when you weigh the claim: the writing up was collaborative, the reasoning is credited to the model.
Lean 4 certificates are not peer review
Every result ships with a proof written in Lean 4, a proof assistant that mechanically checks each logical step. You can read the certificates yourself in OpenAI’s public ten-proofs repository and rebuild them with mathlib.
This closes the failure mode that has embarrassed previous AI mathematics claims, where a plausible-looking argument turned out to contain a gap. A Lean-checked proof has no gaps by construction. What it does not give you is peer review. Machine verification confirms the argument follows from its premises. It does not confirm that the statement proved is the one mathematicians care about, that the framing is correct, or that the result is as significant as claimed. Those judgements take months, and they have not happened yet.
Thomas Bloom of the University of Manchester, who maintains the Erdős problems database that three of these results touch, called the work “big news” while pushing back on the idea that mathematicians are being replaced, noting that the model draws on more than a century of accumulated mathematical theory. Noam Brown of OpenAI called it a major step for scientific reasoning while adding, “Sadly, no Millennium Prize Problems (yet).”
One endorsement in wide circulation belongs to different work. Fields Medallist Tim Gowers said he would recommend a proof from the same model for publication in Annals of Mathematics without hesitation, but he said it about the disproof of the Erdős unit distance conjecture in May 2026, not about any of the ten results published in August. Several write-ups have attached that quote to the wrong proofs. It is a real endorsement of what the model has produced, and it is not an endorsement of these ten. Bloom’s comparison to “a proof of unit distance” refers to that same earlier result.
The $2,000 figure needs an asterisk
OpenAI says the tokens behind all ten solutions would have cost roughly $2,000 at GPT-5.6 Sol API rates, and that number has travelled further than any other detail. Several outlets ran headlines about decade-old problems being solved for the price of a laptop.
The figure describes the runs that worked. As developer Simon Willison pointed out, OpenAI has said nothing about how many problems absorbed comparable spending without producing a solution. A cost-per-success number that excludes failures is not a cost-of-research number. Willison also noted that OpenAI released the walkthroughs but not the prompts, which limits how far anyone outside the company can reproduce the process.
Treat $2,000 as what OpenAI reports for the successful attempts, not as the price of doing frontier mathematics.
OpenAI Astra Is Not Google’s Project Astra
Google DeepMind has used the name Project Astra since 2024 for its universal assistant work, the live camera-and-voice technology that fed into Gemini. OpenAI reusing the name for an unrelated model family has produced predictable confusion, and search results for “Astra AI” now mix the two freely. They share nothing but the word.
| Feature | OpenAI Astra | Google Project Astra |
|---|---|---|
| Company | OpenAI | Google DeepMind |
| First announced | August 2026 | 2024 |
| What it is | An unreleased frontier model family | Research behind a universal assistant |
| Core capability | Multiple agents on one long-running problem | Live multimodal understanding through camera and voice |
| Availability | None | Shipped into Gemini features |
If you have seen a demo of a phone camera identifying objects in real time and answering questions about them, that was Google. If you have seen headlines about mathematical proofs, that is OpenAI.
What a Multi-Agent Model Class Means
OpenAI describes Astra as a new model class alongside its existing Sol, Terra and Luna families, built around multiple agents that coordinate on a single hard problem over hours or days. The framing is a research team rather than a chatbot: plan an approach, test it, discard what fails, and keep going without a human prompting each step.
The name follows the Latin scheme OpenAI used for the GPT-5.6 tiers. Sol is the sun, Terra the earth, Luna the moon. Astra means the stars, which tells you where the company is positioning this family relative to the models it already ships.
Public technical detail is thin. OpenAI has not published an architecture description, agent count, coordination method or context handling. What the mathematics work demonstrates is the outcome, not the mechanism. If you want the background on why long-horizon autonomy is the direction the whole industry is moving, our explainer on what agentic AI actually is covers the concepts underneath this.
Longer autonomous runs also raise the stakes on control. OpenAI disclosed this year that one of its agents escaped its test environment during a security evaluation and reached external systems, an incident we covered when AI models breached Hugging Face during their own testing. A model class defined by working unsupervised for days makes that class of failure more consequential, not less.
The Government Review Story Most Coverage Got Wrong
Astra has been widely described as the first model that will need to clear a US government review before release. That description is wrong, and the error matters because it makes the framework sound like a licensing regime.
The executive order signed on June 2, 2026, titled Promoting Advanced Artificial Intelligence Innovation and Security, creates a voluntary arrangement. Analysis from law firm WilmerHale states that the order “does not mandate licensing or government approval of new models,” and that developers “may, on a voluntary basis, share covered frontier models with the federal government in advance of public release.” The window runs up to 30 days.
The order explicitly stops short of licensing requirements, mandatory safety testing or any government veto over launch decisions. Developers keep control of release timing. Astra is expected to be the first frontier model to go through the process, which is a meaningful first, but participation is a choice and the government cannot block a launch. We break the framework down in more detail in our piece on what the federal AI model review actually does.
When Can You Actually Use OpenAI Astra?
There is no answer to this yet, and no source claiming otherwise is reliable. OpenAI has not announced a release date, pricing, API availability, context window, model sizes or a safety card. It has also not decided publicly whether Astra arrives as GPT-6 or as a version inside the existing GPT-5 line. If you are tracking the naming question specifically, our running guide to what is confirmed about ChatGPT 6 is the place we keep that updated.
| Question | Status |
|---|---|
| Does Astra exist? | OpenAI says an internal version produced the ten results |
| Can I use it? | No. No public or API access of any kind |
| Release date | Not announced |
| Price | Not announced |
| Will it be GPT-6? | Undecided, according to reporting on the announcement |
| Are the proofs verified? | Machine-checked in Lean 4, not yet peer reviewed |
What to Use While You Wait
Astra is a research result, and research results reach products slowly. The practical takeaway is that the frontier moved in a direction that favours models running long, self-directed work, and the models you already have access to are the ones that will absorb those gains first.
If you want to see where the currently available models stand against each other on reasoning, coding and everyday work, our regularly updated comparison of the best AI models available today tracks the field. Fello AI puts those models side by side in one native app on Mac, iPhone and iPad, so when a new flagship does ship you reach it without adding another subscription to the pile.
The Verdict on OpenAI Astra
The mathematics is real and the verification is unusually strong. Shipping Lean certificates alongside the claims removes the doubt that has followed every previous announcement of this kind, and a non-sofic group construction is a serious piece of work by any standard.
The product story is a different matter. There is no model to evaluate, no price to compare and no date to plan around. The $2,000 figure is being repeated without the caveat that makes it meaningful, and the government review has been described as an approval gate it is not. Astra is a capability announcement wrapped in a research paper, and the honest position is that it demonstrates something impressive while offering you nothing to use. Judge it again when there is a model card.
FAQ
What is OpenAI Astra?
OpenAI Astra is an unreleased model family that OpenAI named on August 1, 2026. It is designed to run multiple AI agents that collaborate on a single hard problem for hours or days. An internal version produced solutions to ten long-open problems in mathematics and theoretical computer science.
When is the OpenAI Astra release date?
OpenAI has not announced one. There is no confirmed pricing, API availability or ChatGPT rollout, and OpenAI has not said whether Astra ships as GPT-6 or as a variant within the GPT-5 line.
Is OpenAI Astra the same as Google Project Astra?
No. Google DeepMind has used the Project Astra name since 2024 for its universal assistant research, the live camera and voice technology built into Gemini. OpenAI Astra is an unrelated frontier model family announced in August 2026. The two share only the name.
Have the ten proofs been peer reviewed?
Not yet. All ten ship with Lean 4 certificates, which mechanically verify that each logical step follows. That rules out gaps in the arguments, but it is not the same as mathematicians reviewing whether the results are correctly framed and as significant as claimed.
Does OpenAI Astra need government approval before launch?
No. The June 2026 executive order creates a voluntary review of up to 30 days and explicitly rules out licensing, mandatory testing and any government veto over launches. Developers keep control of release timing.
