OpenAI’s Astra model solved ten open mathematics problems on 1 August 2026, each unsolved for at least a decade, for roughly $2,000 in total compute. The headline result is the first-ever explicit construction of a non-sofic group — a problem open since 1999. All proofs are formalised in Lean 4 with a zero “sorry” count, meaning every step is machine-verified. The structural question the result poses is not whether Astra is impressive. It is what happens to the economics of mathematical research when the marginal cost of a decade-old proof drops to $2,000. [Established on the reported facts; Assessed on structural implications.]
1. What Astra Did
On 1 August 2026, OpenAI announced that an internal, unreleased version of its next major model — named Astra — had solved ten open problems across mathematics and theoretical computer science. [Established — SiliconANGLE, “OpenAI’s Astra solves 10 long-open math problems and publishes the proofs,” 2 August 2026; The Next Web, “OpenAI says its next model, Astra, has solved ten open problems in mathematics,” 2 August 2026; Forbes, “OpenAI’s Astra Solved Decades-Old Math Problems For $2,000,” 3 August 2026.] The problems had been open for at least a decade each. The total compute cost was approximately $2,000. Every proof is formalised in Lean 4, published on GitHub under an Apache 2.0 licence, and carries a “sorry” count of zero — meaning no step in any proof has been asserted without machine verification. [Established — SiliconANGLE, 2 August 2026; DataCamp, “OpenAI’s New Model, Astra, Has Solved Ten Open Math Problems,” citing GitHub repository.]
The headline result is the first-ever explicit construction of a non-sofic group. The question of whether non-sofic groups exist has been open in abstract algebra and group theory since Misha Gromov proposed sofic groups in 1999. No human mathematician had resolved it in twenty-seven years. Astra produced a Lean 4 certificate that constructs one. [Established — The Next Web, 2 August 2026; Quartz, “OpenAI Astra model solves 10 open math problems,” citing Lean 4 repository.]
Other results include a disproof of Connes’s rigidity conjecture — constructing infinitely many non-isomorphic groups with property (T) that share the same von Neumann algebra — and resolutions of two longstanding Erdös problems in extremal graph theory. [Established — SiliconANGLE, 2 August 2026.]
Fields Medal winner Timothy Gowers, speaking to the press, said he would recommend one of the proof family for publication in a top mathematics journal without hesitation. [Established — Forbes, “OpenAI’s Astra Solved Decades-Old Math Problems For $2,000,” 3 August 2026.]
2. Why Lean 4 Matters: The Verification Problem Is Solved
Prior AI mathematics claims have been disputed on a consistent ground: the model may have produced a plausible-looking argument that contains an undetected error. Human mathematics has centuries of experience with proofs that looked correct and were not. The standard rebuttal to any AI proof claim has been: show us the formal verification.
Lean 4 is a proof assistant and programming language designed to produce machine-checkable formal mathematics. A Lean 4 proof is not a natural-language argument that a human evaluates for plausibility; it is a certificate that a theorem-proving engine verifies step by step. A zero “sorry” count means no step has been asserted without being verified — the Lean kernel has checked every inference. [Established — Lean 4 documentation, leanprover.github.io. Tier 1 primary source.]
The practical consequence: anyone can download Astra’s proofs from the GitHub repository, run them through the Lean 4 checker, and independently verify their correctness. No trust in OpenAI is required. The correctness of the result is separable from the credibility of the source. [Assessed — standard formal methods inference; labeled as assessment.]
This is not a marginal point. The history of AI-mathematics interaction has been littered with impressive-sounding claims that dissolved under scrutiny. The Lean 4 certificates make scrutiny straightforward. Gowers’s willingness to recommend the proof for a top journal is premised on this: a verified formal proof is not an AI claim — it is a mathematical fact.
3. The Cost Number Is the Structural Story
The number that the mainstream coverage has largely treated as a curiosity — approximately $2,000 in compute — is the structural story.
A problem open since 1999 means that the combined effort of the mathematics community over twenty-seven years produced no resolution. That effort represents thousands of researcher-years, doctoral programmes, grant cycles, and conference presentations. The marginal cost of all that human effort, allocated to non-sofic groups, is not calculable — but it is orders of magnitude larger than $2,000.
The question the cost number poses is not “is Astra impressive.” It is: what happens to the research allocation logic of mathematics departments when a long-open problem can be resolved for $2,000 of compute?
The first-order answer is straightforward: problems that were open because of human bandwidth constraints — too hard to be solved quickly, not hard enough to attract sustained top-tier attention — become tractable. The second-order question is harder: what happens to the human researchers whose career contribution was to maintain pressure on exactly those problems? And the third-order question is harder still: are there categories of mathematical difficulty that resist this approach, or is the difficulty taxonomy itself about to be rewritten? [Assessed — structural inference; labeled as analysis, not established claim.]
The Leadsman covered the US physics exodus in Sounding No. 6 — a brain drain that Europe is capturing as institutional capital. [Established — The Leadsman, “The Pipeline Transfer,” Sounding No. 6, 7 August 2026.] Astra adds a dimension to that analysis: the institutions that are gaining human mathematical talent may simultaneously be arriving at a moment when the marginal value of that talent, for a specific class of problems, is being compressed by machine proof capability. The two stories are not independent.
4. What Astra Is Not
Astra has no release date, no pricing, and must pass a US government security review before any public rollout. [Established — SiliconANGLE, 2 August 2026; The Next Web, 2 August 2026.] The ten proofs were produced by an internal version of the model — not a product available to researchers or institutions. The gap between a capable internal model and a deployable research tool is substantial in practice: access constraints, safety review, cost structures, and interface design all intervene between a model’s capability and its utility to the broader mathematical community.
The steel-man for scepticism: a model that resolves specific open problems under controlled conditions, with a curated problem set and engineering resources behind the proof-generation pipeline, is not the same as a general-purpose mathematical research assistant. The $2,000 cost reflects OpenAI’s infrastructure and access to Astra’s capabilities; the cost for an external researcher using a future API would be different and unknown. [Assessed — standard capability-versus-access analysis; labeled as assessment.]
What is established: a model has produced machine-verified formal proofs of decade-old open problems for $2,000, and a Fields Medal winner has endorsed one for publication. The structural reorientation this implies for mathematical research economics is real even if the timeline and access questions remain open.
Prediction: At least three of the ten Astra proofs — including the non-sofic group construction — are accepted for publication in peer-reviewed mathematics journals within 24 months, with the formal Lean 4 certificates cited as the basis for the editorial decision.
Confidence: Assessed ~70%. Gowers’s public endorsement establishes a credibility signal; the machine-verifiable format removes the standard objection to AI-derived proofs. The principal failure mode: journal editorial policies prove slower to adapt than the mathematical community’s acceptance of the results.
Resolution: August 2028. Verify: major mathematics journals (Annals of Mathematics, Inventiones Mathematicae) for publication of Astra-attributed proofs.
Bottom line: OpenAI’s Astra solved ten problems the mathematics community had been unable to resolve for a combined span of decades — for $2,000. The proofs are machine-verifiable. A Fields Medal winner has endorsed one for publication. The cost number is not a curiosity. When the marginal cost of a twenty-seven-year-open proof drops to $2,000, the allocation logic of mathematical research changes. The question is not whether this happened. The question is what it means for the researchers, institutions, and funding bodies whose strategic calculus assumed it could not.