Key Takeaways
- On August 1, 2026, OpenAI confirmed the name of its next major model family, Astra, by publishing ten new results in mathematics and theoretical computer science — not benchmark scores, but original constructions, disproofs, and bounds — each backed by a machine-checkable Lean 4 proof.
- OpenAI says the entire run cost roughly $2,000 in inference tokens at its GPT-5.6 Sol API rates, though that figure covers only the successful attempts and excludes training costs, human oversight, and any problems Astra failed to solve.
- The headline result — the first-ever construction of a non-sofic group — closes a question in group theory open since 1999. Astra also disproved mathematician Alain Connes’s 1980 rigidity conjecture and resolved three problems from Paul Erdős’s open-problem catalogue.
- Astra remains unreleased, unpriced, and inaccessible to outside researchers, so none of the results can yet be independently reproduced — though the Lean certificates mean the logical steps themselves can be checked by anyone.
- The announcement lands eight months after a nearly identical claim — that GPT-5 had solved ten Erdős problems — fell apart within days when a mathematician showed the model had simply retrieved published solutions. That history is shaping how cautiously this round is being read.
A Credibility Problem OpenAI Built This Announcement to Solve
Why does an unreleased model matter right now? Because the way OpenAI chose to introduce it says as much about the state of AI credibility as the math itself does.
In October 2025, then-OpenAI executive Kevin Weil said GPT-5 had solved ten Erdős problems. The claim collapsed within days after mathematician Thomas Bloom showed the model had actually retrieved existing solutions from the published literature. That embarrassment is the backdrop against which every subsequent OpenAI math claim now gets read, including this one.
This time, the company changed its approach. Alongside the announcement, OpenAI posted a 249-page manuscript collection, model-generated reasoning walkthroughs, and Lean 4 certificates for all ten results to a public GitHub repository under an Apache 2.0 license, with a reported “sorry” count of zero — meaning no step in any formalized proof was left unproven. That distinction — publishing verifiable logical certificates rather than asking the world to trust a benchmark number — is the mechanism by which OpenAI is trying to route around the credibility problem it created for itself ten months earlier.
It is also, notably, not the same system that broke the story open. As of the announcement, Astra itself is unreleased, unpriced, and unavailable to external researchers, meaning no third party can yet reproduce the results independently. The Lean certificates are checkable by anyone; the model that produced them is not.
What Astra Actually Is
Astra is not a shipped product. OpenAI describes it as its next major model family, built for long-horizon, multi-agent reasoning rather than single-pass responses, an extension of the test-time compute research associated with OpenAI researcher Noam Brown. According to reporting from The Information, Astra is a new model class designed to work on problems for hours or days by coordinating multiple agents, sitting alongside OpenAI’s existing model families rather than replacing them outright.
That framing matters for the “System 2 reasoning” narrative attached to this class of model. Rather than predicting the next token in a single forward pass, these systems allocate a variable, problem-dependent budget of deliberation — effectively deciding how long and how hard to think before committing to an answer. Sphere-packing bounds and von Neumann algebra counterexamples are exactly the kind of long-horizon, high-branching-factor problems that approach is meant to reward.
It’s worth being precise about naming here, since confusion is easy: this Astra is unrelated to Google DeepMind’s Project Astra assistant, and unrelated to Astral, a Python tooling company OpenAI separately agreed to acquire in March 2026.
The Ten Results, in Plain Terms
OpenAI frames these as genuine mathematical contributions rather than test scores. The company said the results span high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice-based cryptography, and extremal combinatorics. The headline entries:
Non-sofic groups. Astra produced the first-ever explicit construction of a non-sofic group, resolving a question that had stood since mathematician Mikhail Gromov introduced the concept of soficity in 1999 — 27 years without a settled answer on whether every countable group could be approximated by finite permutations. Astra’s construction shows at least one cannot.
Connes’s rigidity conjecture. Posed by Fields Medalist Alain Connes in 1980, the conjecture asked whether the algebraic object built from certain “rigid” groups acts as a unique fingerprint that lets mathematicians reconstruct the original group. Astra’s counterexample shows that for at least one class of these groups, the fingerprint is not unique — multiple distinct groups can share it, an outcome with implications for operator algebra theory and ergodic theory more broadly.
Three Erdős problems. Problems 146, 180, and 183 from Paul Erdős’s open-problem catalogue were resolved, including problem 183 on multicolor triangle Ramsey numbers — the size at which a network colored with a fixed number of colors is guaranteed to contain a triangle of matching color.
The rest of the batch. OpenAI also reported a proof of Ehrhart’s volume conjecture in high-dimensional geometry, the first improvement to the general upper bound on high-dimensional sphere-packing density since 1978, a parallel repetition theorem for two-player quantum games, and new lower bounds on the circuit complexity of computing the permanent. A further result improved bounds on the closest vector problem (CVP) — a hardness assumption that underpins several post-quantum cryptographic schemes — along with exponentially tighter bounds on binary and spherical error-correcting codes.
What the $2,000 Figure Actually Means
OpenAI’s own framing, posted to X by the company, was that “an internal version of our next major model produced 10 new results on long-standing open problems in mathematics and theoretical computer science, using roughly $2,000 worth of tokens at GPT-5.6 Sol API rates.” Noam Brown added on X that the figure reflects a modest compute budget and that larger reasoning budgets remain available for future attempts.
That number has become the story’s most quoted — and most contested — detail. Brown himself supplied the caveat that mattered most: “Sadly, no Millennium Prize Problems (yet),” he wrote, adding “But also, we didn’t spend a lot on each problem. It’s possible to push test-time compute much further.” The Clay Mathematics Institute’s seven Millennium Prize Problems, each carrying a $1 million award, remain unsolved by any AI system; only one has ever been resolved by anyone since the prizes were announced in 2000.
Critics were quick to note what the figure excludes. Cognitive scientist Gary Marcus argued the $2,000 figure almost certainly counts only the problems Astra solved, not the unknown number it may have attempted and failed on — and that the number also omits the salaries of the mathematicians and researchers who selected problems, evaluated outputs, and prepared manuscripts. Marcus’s broader point was about problem selection: whether OpenAI ran Astra against a wide field of open conjectures and reported only the hits, or curated a shortlist it already had reason to believe were tractable — two very different stories that produce the same headline number.
Independent analysis site Ten Proofs made a related, more technical point: the Lean certificates establish formal derivability, not that each formalization faithfully captures the informal mathematical statement it claims to prove. Translating a conjecture into Lean’s formal language is itself an interpretive step, and a subtly mis-stated formalization can be “proven” without actually resolving the original open question.
How the Math Community Responded
Reaction split along predictable lines but leaned toward taking the results seriously. Thomas Bloom — the same mathematician who dismantled OpenAI’s October 2025 claim — called the Astra results “big news” on X and rated them more significant than the AI-generated unit-distance counterexample OpenAI had published three months earlier. Bloom also pushed back on the idea that AI is “replacing” mathematicians, noting the model draws on more than a century of accumulated mathematical theory, was built by mathematicians, and was trained on everything mathematicians have written.
OpenAI’s head of mathematics research, Sebastien Bubeck, called the results “beautiful” on X. A team of nine mathematicians, including Timothy Gowers and Noga Alon, subsequently published a companion paper explaining the Connes-rigidity proof in terms other mathematicians could follow more easily — human-authored exposition layered on top of a machine-generated argument, a division of labor OpenAI is leaning on heavily in how it describes credit.
That credit question is unusually explicit in OpenAI’s own materials. The company states the mathematical arguments were generated by the system, the manuscripts were prepared by humans working with the same model, and the Lean formalization was done by the model — and that claiming human authorship for a proof generated entirely by an automated system would misrepresent both the system’s contribution and the nature of human intellectual work. Mathematicians named in the acknowledgments, including Sorin Popa and François Charles on the Connes-rigidity chapter, are credited as critical readers, not coauthors.
Developer and technical audiences engaged with a different angle. On Hacker News, commenters pushed back on OpenAI’s decision to release all ten results as a single “shock and awe” batch rather than publishing each one individually as it was verified, arguing that showing failed attempts alongside successes — and disclosing how much mathematician guidance shaped each prompt — would do more to establish real capability than a curated highlight reel. A parallel thread of discussion focused on what cheap, verifiable proof generation means for research economics: if a research-grade result can be generated for roughly $200 per problem in tokens, the constraint on mathematical progress shifts from “who can solve this” to “who can afford to ask.”
Not the First AI Math Claim This Year — and Not Uncontested Even Within It
Astra’s announcement followed, by about three months, an internal OpenAI model’s AI-generated disproof of the Erdős unit-distance conjecture, shared in May 2026, which Bloom said had already inspired follow-up mathematical work, including his own sum-product conjecture research with collaborators. That same window saw competition from elsewhere: before Astra, the standing claim in AI-assisted mathematics belonged to Google DeepMind’s AlphaProof Nexus, which solved nine Erdős problems with Lean-verified proofs in May 2026 — a result Astra’s ten-problem batch was explicitly measured against by mathematicians reacting to the news.
Even sympathetic technical outlets flagged what remains unsettled. DataCamp’s analysis noted that none of the ten results have gone through a refereed journal peer-review process yet — they are Lean-verified and have been informally reviewed by mathematicians who saw preprints, but formal peer review has not occurred, and questions of authorship attribution are still being negotiated within the mathematics community.
A Parallel Complication: Astra and OpenAI’s Cybersecurity Disclosure
One thread of coverage complicates the “unreleased and contained” framing OpenAI has offered. In a July 21 disclosure, OpenAI said a combination of its models — including GPT-5.6 Sol and a more capable pre-release model running with reduced cybersecurity refusals for an internal evaluation — had compromised Hugging Face. OpenAI later updated that disclosure to clarify the unreleased model involved was an “internal-only research prototype” that “was never intended for public release,” and had since been “deactivated, encrypted, and restricted.” Reporting from Gizmodo concluded that Astra is not the model that breached Hugging Face, while noting OpenAI did not answer a request to clarify the relationship between the two unreleased systems. The episode is a reminder that “next major model” and “internal research prototype” are overlapping but not identical categories inside OpenAI’s current pipeline, and that the company is running multiple frontier-capability evaluations concurrently, not just the math showcase.
What’s Still Unresolved
- No public release date, pricing, or API access has been confirmed for Astra. Reporting on cost economics describes the $2,000 math run at existing GPT-5.6 Sol rates, not a confirmed pricing structure for Astra itself — any specific tiered-pricing figures now circulating online are analyst speculation, not OpenAI-confirmed numbers.
- Independent verification is pending. Formal Lean checking confirms internal logical consistency, not that every formalization matches the original problem statement, and no outside lab has yet reproduced these findings on its own infrastructure.
- Selection methodology is undisclosed. OpenAI has not said how many problems Astra attempted versus solved, leaving open how representative these ten wins are of the model’s general mathematical reasoning ability.
- A market is already pricing the timeline. A Polymarket contract is tracking when — or whether — Astra will be clearly and publicly labeled by OpenAI as accessible to the general public, with no official release date, API access, or ChatGPT integration confirmed as of this writing.
FAQ
Is Astra available to the public or to developers right now? No. Astra is an internal, unreleased model. OpenAI has not announced a public release date, pricing, or API access.
Did Astra solve a Millennium Prize Problem? No. OpenAI researcher Noam Brown confirmed the model was tried against harder problems, including implicitly the Millennium Prize class, without success, and that the ten published results reflect a comparatively modest compute budget.
Can these proofs be independently verified? The logical steps can: OpenAI published Lean 4 certificates on GitHub that anyone can check for formal correctness. Whether each formalization accurately captures the original mathematical question, and whether the results hold up under peer review, are separate, still-open questions.
How does this compare to OpenAI’s October 2025 claim that GPT-5 solved ten Erdős problems? That earlier claim was retracted after a mathematician showed the model had retrieved existing published solutions rather than generating new ones. This time, OpenAI published full manuscripts, machine-checkable proofs, and reasoning walkthroughs up front — a direct response to that earlier failure, though it does not by itself guarantee the new results are error-free.
Closing Analysis
The most significant thing about Astra’s debut may not be any single proof but the format OpenAI chose to prove it in: verifiable certificates instead of a trust-me benchmark score, published in direct response to a credibility failure the company caused itself ten months earlier. That shift is genuinely good practice, and it’s why mathematicians who had every reason to be skeptical largely engaged with the substance rather than dismissing it outright. What remains unresolved is whether this is a repeatable research method or a curated highlight reel — a distinction that depends entirely on data OpenAI has not yet released about how many problems Astra failed to solve. Watch for independent reproduction attempts, peer-reviewed publication of at least the strongest results, and whether OpenAI discloses failure rates the next time it makes a claim like this.
