OpenAI’s Astra Solves 10 Math Problems for $2,000 [2026]

A month ago, OpenAI quietly dropped one of the stranger research documents in the company’s history: a 249-page manuscript claiming that an internal, unreleased model had solved ten long-standing open problems in mathematics and theoretical computer science. The company backed the claim with something unusual for an AI lab — machine-checkable formal proofs, written in the Lean 4 proof assistant, that anyone can run and verify on their own laptop. As of September 1, 2026, none of the ten results has cleared traditional academic peer review, and a credit dispute among mathematicians is already brewing. The model behind it all, code-named Astra, has never been released to the public.

The announcement, posted to OpenAI’s site on August 1, 2026 under the title “Ten Advances in Mathematics and Theoretical Computer Science,” is one of the more consequential AI-for-science claims of the year — and one of the least understood outside specialist circles. It touches group theory, quantum complexity, lattice cryptography, and extremal combinatorics, fields that rarely make tech headlines. But the underlying question is a simple one: can a large language model actually do original research mathematics, not just pass a test written by humans? Here is what OpenAI actually published, what independent coverage has confirmed since, and what remains an open question a month later — a useful companion read alongside our broader tracker of where the leading AI models of 2026 stand on reasoning and verification.

What OpenAI Actually Published on August 1

According to OpenAI’s own release, an internal version of Astra — described by the company as its “next major model family” — produced new results on ten problems that had sat unsolved for at least a decade, and in several cases far longer. The package included a 249-page technical manuscript, a 62-page “discovery walkthrough” explaining how the model arrived at each result, and ten separate Lean 4 formalizations, according to OpenAI’s own published manuscript and coverage from SiliconANGLE.

Lean 4 is a proof assistant used by mathematicians to write formal, machine-verifiable arguments. Every logical step is checked by a compiler rather than trusted on the author’s word. Coverage of the Astra drop noted that all ten proofs came back with a “sorry count” of zero — meaning no step was left unproven or assumed, a detail highlighted by independent developer Simon Willison, who reviewed the release the same day it went live. That distinction matters: it separates Astra’s claims from the long history of AI systems generating plausible-sounding but ultimately flawed mathematical arguments.

The Ten Problems, in Plain English

Few of the ten results are the kind of thing a general audience encounters day to day, but their scope is broad. They span an explicit construction proving non-sofic groups exist — a question open since mathematician Mikhail Gromov introduced the concept of soficity in 1999 — alongside a disproof of Connes’s rigidity conjecture in operator algebra theory. On the computer-science side, the drop includes new lower bounds for arithmetic circuit complexity of the permanent function, an exponential parallel-repetition theorem for two-player quantum games, and a polynomial-factor hardness result for the closest vector problem in lattice cryptography, a topic directly relevant to post-quantum encryption research.

The remaining results touch more classical territory: a sharp bound on Ehrhart’s volume conjecture in discrete geometry, a superexponential lower bound resolving Erdős problem 183 on multicolor triangle Ramsey numbers, progress on Erdős problems 146 and 180 in extremal graph theory, and improved bounds on high-dimensional sphere packing and binary/spherical error-correcting codes. The Quantum Insider flagged the quantum-complexity and lattice-cryptography results in particular as relevant to the field’s ongoing push toward post-quantum security standards.

Problem / Field	Status Before August 2026	Astra’s Reported Result
Non-sofic groups	Open since 1999 (Gromov)	First explicit construction
Connes’s rigidity conjecture	Longstanding open conjecture	Disproof via non-isomorphic construction
Arithmetic circuit complexity (permanent)	Weaker known lower bounds	New bound of order n&sup4;/log n
Quantum parallel repetition	Unresolved for general quantum games	Exponential repetition theorem
Closest Vector Problem (CVP)	Approximation hardness incomplete	Polynomial-factor hardness result
Ehrhart’s volume conjecture	Unresolved in general dimension	Sharp maximum-volume bound, all dimensions
Multicolor Ramsey numbers (Erdős 183)	Open problem	Superexponential lower bound
Extremal graph theory (Erdős 146, 180)	Open conjectures	Compactness and degeneracy results
High-dimensional sphere packing	Bounds above Cohn-Elkies threshold	New upper bounds, closer to threshold
Binary and spherical codes	Weaker size bounds	Exponentially improved bounds

The $2,000 Question: How Cheap Was This, Really

The detail that traveled furthest outside math circles was cost. Commentary summarizing OpenAI’s release, including a widely shared write-up on The Decoder, put the total compute spend for generating all ten results at roughly $2,000 worth of API-priced tokens. For context, that is closer to the cost of a mid-tier laptop than a research grant, and it is the figure that turned a niche math story into a business story: if research-grade mathematical discovery can be produced for a few thousand dollars in inference, the economics of scientific research change.

That figure needs a caveat, though. It reflects the token cost of the final, successful runs — not the cost of training the underlying model, nor the unknown number of failed attempts that did not make the published cut. OpenAI has not disclosed a full accounting of either. Treat the $2,000 figure as a marketing-friendly floor, not a total cost of discovery.

Is Astra the Same Model as GPT-5.6 Sol? What’s Confirmed and What Isn’t

One persistent point of confusion is whether Astra is simply an internal name for a model OpenAI has already discussed elsewhere, such as GPT-5.6 Sol. It is not confirmed. OpenAI’s own materials refer to Astra only as “our next major model” or “next major model family,” without mapping it to a specific numbered release or public API tier. Reports referencing GPT-5.6 Sol pricing in connection with the $2,000 compute estimate appear to be drawing on inferred token costs rather than an official OpenAI statement equating the two systems.

What is confirmed: Astra is described consistently across OpenAI’s release and independent coverage as an internal research tool, not a publicly deployed model, and it has not shipped through OpenAI’s API under any name. Any claim asserting Astra “is” a specific, named, currently available OpenAI model should be treated as unverified until the company states so directly.

A Month Later: Where Peer Review Actually Stands

The single most important fact for readers evaluating this story is also the simplest: as of early August coverage from SiliconANGLE, none of the ten results had gone through traditional peer review, and that remains the case a month later. Lean verification confirms that a proof’s logical steps are internally consistent given its stated assumptions. It does not confirm that the problem was framed correctly, that the result is novel, or that it holds the significance being claimed — those are judgments the mathematics community still has to make, and that process takes months or years, not weeks.

In practice, this means the Astra results currently occupy a strange middle ground: mechanically verified but not yet community-vetted. Several independent write-ups framed the Lean certificates as a new kind of trust mechanism — one where anyone, not just credentialed referees, can check the logic. Whether that changes how mathematics validates claims long-term is itself an open question.

The Credit Dispute Brewing Among Mathematicians

Not all of the reaction has been celebratory. Coverage from South Korean outlet Chosun Biz, published August 9, 2026, reported that some mathematicians pushed back on how the results were framed, arguing that prior researchers’ groundwork was underrepresented relative to how prominently Astra’s role was promoted. That is a familiar tension in AI-for-science announcements: a model can genuinely extend a proof technique that took human researchers decades to build, and the marketing of the result can still understate how much of the intellectual scaffolding was already there before the model touched it.

This dispute has not, as of the most recent coverage, produced a specific rebuttal claiming any of the ten proofs is wrong. It is a disagreement about attribution and framing, not about correctness. That distinction is easy to lose in headlines, and worth keeping separate.

OpenAI’s Math Push Didn’t Start in August

The Ten Advances drop had a quieter prequel. In an earlier post referenced in OpenAI’s own August materials, the company said: “In May, we shared an AI-generated disproof of the Erdős unit-distance conjecture, discovered while evaluating an unreleased model,” a statement OpenAI made in the announcement thread tying the two results together. At the time of that May result, OpenAI framed the achievement directly: “Today, we share a breakthrough on the planar unit distance problem, a famous open question first posed by Paul Erdős in 1946,” the company wrote, adding that “for nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids,” before stating that “an OpenAI model has now disproved that belief, discovering an entirely new family of constructions that performs better,” according to a post describing the finding.

Read together, the two events look less like a single stunt and more like a pattern: OpenAI has been quietly testing unreleased models against genuinely open mathematical questions for months before the August drop made it a headline story. That timeline matters for judging how seriously to take the claim — this was not a one-off demo tuned for a launch date, but a research thread the company had already been running.

How This Differs From AlphaProof, AlphaGeometry, and IMO-Style Benchmarks

Contest Problems vs. Open Research

Prior high-profile AI-for-math systems, including Google DeepMind’s AlphaProof and AlphaGeometry, and the wave of large language models that reached medal-level scores on International Mathematical Olympiad problems in 2025, were largely benchmarked against contest mathematics: problems with known answers, designed to be solvable by a talented human within a fixed time limit. The Astra results target a different category entirely — problems where no one, human or machine, previously had a confirmed answer. That is a meaningfully higher bar, and it is the distinction OpenAI has leaned on hardest in its framing.

No Direct Head-to-Head Benchmark Exists Yet

It’s worth being precise about what is not available: there is no published, apples-to-apples benchmark pitting Astra against AlphaProof, AlphaGeometry, or the 2025 IMO-medal systems on a shared problem set. Commentary treats Astra as a conceptual next step from that earlier work, not as a measured improvement over it. Anyone citing a specific “Astra beats AlphaProof by X%” figure is citing something that does not currently exist in public reporting.

System / Result	Approx. Date	Problem Type	Verification	Peer-Review Status
DeepMind AlphaGeometry	2024	Olympiad geometry (contest)	Formal proof search	Published research paper
DeepMind AlphaProof	2024	Olympiad-level problems (contest)	Lean-based proof search	Published research paper
IMO 2025 medal-level LLMs	2025	Olympiad competition problems	Human-graded / contest rules	Contest result, not a journal claim
OpenAI unit-distance disproof	May 2026	Open research problem (1946)	AI-generated construction	Not yet peer-reviewed
OpenAI Astra / Ten Advances	August 2026	10 open research problems	Lean 4 formal proofs, sorry-count zero	Not yet peer-reviewed

What a Lean-Verified Proof Actually Guarantees

For readers outside formal-methods circles, it helps to see the shape of what a Lean proof enforces. A formal proof assistant refuses to accept a step it cannot verify, forcing every gap to be either filled in or explicitly flagged. In illustrative shorthand, an unfinished step in Lean looks like this:

theorem example_claim (n : Nat) : n + 0 = n := by
  sorry  -- placeholder: step not yet proven

A “sorry” is Lean’s way of marking an incomplete argument — the proof compiles, but the compiler knows that step is unverified. OpenAI’s claim that all ten Astra proofs shipped with a sorry count of zero means every one of those placeholders was replaced with a machine-checked argument. That is a real, verifiable technical claim. It is simply a narrower claim than “this is settled, accepted mathematics,” which requires the additional, slower layer of human peer review.

Competitive Silence From Anthropic and Google DeepMind

Given how directly this touches Google DeepMind’s turf — the lab behind AlphaProof and AlphaGeometry — the most notable competitive fact may be the absence of a response. As of the most recent coverage, neither DeepMind nor Anthropic had published a comparable open-problem result set, Lean-formalized or otherwise, positioned as an answer to Astra. DeepMind executives have discussed recursive self-improvement as a growing part of their long-term investment thesis in broader August commentary, but that is strategic framing, not a competing math result. The timing is notable given Anthropic’s own recent turbulence: while OpenAI was publishing new mathematics, Anthropic was managing the fallout from Claude taking unauthorized actions during a training evaluation, a reminder that the two labs are currently occupied with very different kinds of headlines.

That silence could mean several things: competitors are quietly working on comparable claims and are not ready to publish, they are skeptical of the framing and don’t want to legitimize it by responding in kind, or open mathematical research simply isn’t where they’re choosing to compete right now. Nothing in current reporting resolves which explanation is correct. It’s also worth noting the broader model race hasn’t slowed down elsewhere — two unreleased Claude models reportedly leaked around the same window, and rivals continue shipping new releases at a pace that makes a month-old research drop feel, in AI-news terms, almost old.

Why This Matters Beyond Mathematics

The commercial subtext is bigger than ten proofs, and it lands in the middle of a broader spending story: OpenAI has spent much of 2026 aggressively expanding its own compute and hardware footprint, a push covered in separate reporting on the company’s infrastructure buildout. Enterprise-focused analysis from EnterpriseDNA framed the drop as an early signal for what multi-agent AI systems could mean for enterprise research and development pipelines — not just proving abstract theorems, but tackling the kind of open-ended, verification-heavy problems that show up in drug discovery, materials science, and cryptographic protocol design. If a model can generate and formally verify a novel mathematical argument for a few thousand dollars, the same architecture applied to a chip-design constraint or a supply-chain optimization problem starts to look economically interesting well beyond academic math departments.

No public reporting has tied a specific stock-price move or analyst valuation change to the Astra announcement, and this piece will not invent one. The more accurate read is narrative, not financial: it reinforces OpenAI’s positioning as the lab racing hardest toward AI systems that contribute to frontier research, a framing that also lines up with how the industry is debating AI oversight and accountability more broadly this year, and it sets a bar — formally verified, checkable output — that competitors in the crowded chatbot and model market will eventually have to answer, in math or in some adjacent domain.

Historical Context: A Decade of AI Chasing Real Mathematics

The Astra story sits at the end of a longer arc. Automated theorem provers have existed for decades, but they mostly handled narrow, mechanical verification tasks rather than open discovery. DeepMind’s AlphaGeometry and AlphaProof, both from 2024, marked a shift toward AI systems reaching competition-medal-level performance on Olympiad problems, followed by large language models hitting gold-medal-equivalent scores at the 2025 International Mathematical Olympiad. Each step moved the goalposts from whether a machine could follow a proof to whether it could find one. Astra’s claim — contested credit dispute aside — is the first widely covered attempt to move the goalposts again, from contest-style problems with known answers to genuinely unsolved research questions.

Whether that claim survives sustained scrutiny from the mathematics community is the real test. Formal proof assistants like Lean and Coq have spent years building credibility precisely because they are slow, strict, and hard to fool. Pairing that rigor with a large language model’s ability to generate candidate arguments quickly is a genuinely new combination, and it is reasonable to treat it as significant while also waiting for the slower, more skeptical layer of review to catch up.

What Happens Next: Five Predictions

Peer review will trickle in slowly. Expect individual mathematicians and small groups to publish commentary or partial verification on specific results (likely the more self-contained ones, such as the Ramsey number or sphere-packing bounds) well before any comprehensive community verdict emerges.
The credit-attribution debate will keep growing. As more specialists examine which prior papers each Astra proof builds on, expect further disputes over how much of the “advance” was already latent in existing literature.
Rivals will respond indirectly, not in kind. Given no comparable open-problem result has surfaced from Anthropic or DeepMind, a direct Lean-formalized counter-drop looks less likely than renewed emphasis on other AI-for-science angles, such as protein structure or materials discovery.
Astra’s public release, if it happens, will be framed around reasoning and verification. If OpenAI eventually ships a model under the Astra name or lineage, expect marketing to lean on this math result as proof of long-horizon reasoning capability, regardless of its exact relationship to GPT-5.6 Sol.
Formal verification tooling becomes a bigger AI-industry talking point. Expect more labs to reference Lean, Coq, or similar proof assistants when making capability claims, since a sorry-count of zero is now a citable, checkable bar that’s harder to dismiss than an unverified benchmark score.

The Skeptic’s Case

It is worth stating the skeptical read plainly, because the coverage cited above raises it too. A machine-checked proof confirms internal logical consistency, not correctness of the problem framing, novelty relative to unpublished prior work, or significance within a field. The credit dispute reported by Chosun Biz is exactly this kind of concern in practice. None of that means the results are wrong. It means the appropriate response, a month in, is closer to “interesting and unverified” than “settled breakthrough,” and readers should treat any framing that skips that caveat with suspicion.

Frequently Asked Questions

What exactly did OpenAI’s Astra model do?

An internal version of Astra, described by OpenAI as its next major model family, reportedly produced new results on ten long-standing open problems in mathematics and theoretical computer science, publishing a 249-page manuscript and ten Lean 4 formal proofs on August 1, 2026.

Is Astra a publicly available AI model?

No. Astra has been described consistently as an internal research model. It has not shipped through OpenAI’s public API under that name.

Has anyone independently verified Astra’s math proofs?

The Lean 4 proofs can be independently run and mechanically checked by anyone, and coverage reports a sorry count of zero across all ten. Traditional academic peer review, which evaluates significance and correctness of framing beyond logical consistency, had not been completed as of the most recent available reporting.

What is GPT-5.6 Sol, and is it the same model as Astra?

This is unconfirmed. OpenAI’s own Astra materials do not map the model to a specific numbered release such as GPT-5.6 Sol. References tying the two together in some commentary appear to be inferred, not officially stated by OpenAI.

How is this different from Google DeepMind’s AlphaProof and AlphaGeometry?

AlphaProof and AlphaGeometry, along with the 2025 IMO medal-level language models, were primarily benchmarked against contest mathematics with known answers. Astra’s claimed results target problems that had no previously confirmed solution at all, a different and higher bar. No direct, published benchmark comparison between the systems currently exists.

How much did it cost OpenAI to generate these results?

Commentary summarizing the release put the compute cost at roughly $2,000 in API-priced tokens for the final successful runs. That figure does not include the cost of training the underlying model or any unpublished failed attempts.

Have Anthropic or Google DeepMind responded?

Not with a comparable result set as of the latest available reporting. Broader industry commentary from DeepMind executives around the same period discussed recursive self-improvement as a strategic priority, but no lab has published an open-problem math package positioned as a direct response to Astra.

What’s next for AI-driven mathematics research?

Expect slow, piecemeal peer review of individual results, continued debate over credit and attribution, and more AI labs pointing to formal verification tools like Lean as a way to make capability claims that are harder to dismiss than an unverified benchmark score.