OpenAI announced on 2 August 2026 that an internal version of its Astra model family solved ten open problems in mathematics, quantum complexity theory, and theoretical computer science — each unsolved for at least ten years. The problems include an explicit construction of a non-sofic group (open since Gromov, 1999), a disproof of Connes’s rigidity conjecture, a proof of Ehrhart’s volume conjecture, and three entries from Paul Erdős’s open problem catalogue. OpenAI released all ten as Lean 4 formal proofs on GitHub, with a stated “sorry count” of zero — meaning no proof step is unverified. Compute cost: approximately $2,000 at GPT-5.6 Sol API rates. The proofs have not yet been peer-reviewed in the traditional sense, and OpenAI has noted that independent scrutiny remains ongoing. The result is the most significant demonstration to date of AI as a primary generator of novel mathematics rather than a tool for verification or search.
1. The Ten Problems: A Reading Guide
OpenAI’s announcement on 2 August 2026 identified ten open problems solved by the Astra model. [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,” August 2026; The Quantum Insider, “OpenAI Says Next-Generation Model Solved 10 Major Open Problems in Quantum Complexity, Mathematics,” 4 August 2026.] The full list spans pure mathematics, quantum complexity, and theoretical computer science. Among the most significant:
Non-sofic group construction. The most celebrated result. Mikhail Gromov introduced the concept of soficity in 1999 as a generalisation of amenability — a property that makes groups tractable for certain analytical methods. Every known group was either proven sofic or not known to be non-sofic. Astra produced an explicit example of a non-sofic group, settling the existence question that has been open for 27 years. [Established — SiliconANGLE, 2 August 2026; The Next Web, August 2026, describing the non-sofic group as “chief among the findings.”]
Connes’s rigidity conjecture disproved. Alain Connes’s conjecture on von Neumann algebras was a foundational claim in functional analysis that had resisted proof for decades. Astra did not prove it; it disproved it — constructing a counterexample. A disproof is, in some respects, harder than a proof: it requires finding a single object that violates a property presumed universal. [Established — The Quantum Insider, 4 August 2026.]
Three Erdős problems. Paul Erdős maintained a catalogue of open problems across combinatorics and number theory, many with prize money attached. Astra resolved three, including problem 183 on multicolour Ramsey numbers. [Established — SiliconANGLE, 2 August 2026; Yahoo Tech/QZ, “OpenAI Astra model solves 10 open math problems for $2,000,” August 2026.]
Ehrhart’s volume conjecture. A problem in combinatorial geometry concerning the relationship between a polytope’s volume and the integer lattice points it contains. [Established — The Quantum Insider, 4 August 2026.]
2. The Lean 4 Proofs: What “Zero Sorry Count” Actually Means
The distinction that separates the Astra announcement from earlier AI mathematical demonstrations is the release format. OpenAI published a 249-page manuscript alongside Lean 4 proof certificates on GitHub under an Apache 2.0 licence. The repository’s “sorry count” stands at zero. [Established — SiliconANGLE, 2 August 2026; The Decoder, “OpenAI announces its ‘next major model’ Astra by dropping ten previously unsolved math solutions,” August 2026.]
In the Lean proof assistant, “sorry” is a keyword that allows a proof step to be marked as accepted without verification — a placeholder that tells the system “trust me here.” A sorry count of zero means that no step has been accepted on trust. Every logical inference in every proof has been checked by the Lean 4 kernel. This is machine-verifiable correctness: not peer review in the traditional sense, but something in certain respects more rigorous, because it does not depend on human readers following a chain of reasoning but on an automated checker verifying each step against the formal rules of the proof system.
What this does not guarantee is mathematical significance. Lean 4 can verify that a proof is formally correct without telling you whether the problem it solves is interesting, whether the construction is elegant, or whether the result connects to other open questions in productive ways. Human mathematical judgment is still required to assess those dimensions. [Assessed — standard interpretation of formal proof verification; the Lean 4 system cannot evaluate mathematical significance, only formal correctness.] OpenAI noted that independent scrutiny remains ongoing, and peer review by the broader mathematics community has not yet been completed. [Established — DataCamp, “OpenAI’s New Model, Astra, Has Solved Ten Open Math Problems,” August 2026: “OpenAI has not yet released detailed proofs or announced peer-reviewed validation.”]
3. The $2,000 Number and What It Reveals
OpenAI stated that the total compute cost for all ten solutions was approximately $2,000 at GPT-5.6 Sol API rates. [Established — Yahoo Tech/QZ, “OpenAI Astra model solves 10 open math problems for $2,000,” August 2026.] This number is striking but requires careful reading.
It is not the total cost of developing Astra, which is many orders of magnitude larger. It is the marginal inference cost of running the model against these ten specific problems — the compute burned to produce the outputs, not to produce the model that generated them. What it does demonstrate is that once a model capable of mathematical research at this level exists, the per-problem cost of deploying it is trivially low compared with the cost of human mathematical research at equivalent depth. A decade of a professional mathematician’s time on a single hard problem costs orders of magnitude more than $2,000 in salary, overhead, and institutional support.
The structural implication is straightforward: if the model generalises across problem types at similar cost, the economics of mathematical research change qualitatively. Priority shifts from which problems to work on toward which problems to ask the model to work on. The bottleneck is no longer compute or human time; it becomes problem selection and result evaluation. [Assessed — analytical inference from the $2,000 compute figure; the generalisability of Astra’s performance across problem types is assessed, not established, pending further public testing.]
4. What This Changes — and What It Does Not
The standard objection to AI mathematical results is that the model has been trained on the existing mathematical literature and may be pattern-matching rather than reasoning. For the Astra results, this objection is partially addressed by the formal proof certification: the Lean 4 kernel does not know whether a result is “elegant” or “expected,” but it does verify that the logical steps are valid. If the proof is valid, the result is correct — regardless of how the model arrived at it.
The residual objection is that Astra may be solving problems that are hard for humans but tractable by exhaustive search in formal proof space — not the same as genuine mathematical creativity. The non-sofic group construction is the hardest counter-argument to that view: constructing an explicit example of an object that has been sought for 27 years, across a large community of specialists, is not obviously a tractable search problem. [Assessed — the difficulty of non-sofic group construction is assessed based on the duration of the open problem and the community that worked on it; the Navigator cannot independently verify the mathematical complexity characterisation without specialist expertise.]
What changes is the research workflow. Professional mathematicians who work in areas adjacent to Astra’s demonstrated capabilities will, rationally, begin using it as a search tool for results they suspect exist but cannot prove. The mathematics community’s peer-review infrastructure — built around human-readable proofs evaluated by domain specialists — will encounter pressure to integrate formal verification into standard practice, because a Lean 4 certified proof does not require a human reader to follow every step.
Bottom line: OpenAI’s Astra has produced the most significant demonstration of AI as a mathematical research tool to date — not because it solved ten problems, but because it released the proofs in a format that is machine-verifiable and publicly auditable. The non-sofic group construction alone, if it survives community scrutiny, represents a result that would be considered major work from a human mathematician. The pending question is what the mathematics community does with it: whether formal proof verification becomes standard practice, and whether the problems Astra cannot yet solve reveal the actual frontier of current AI reasoning capacity. The Astra results are real, certified, and consequential. They are also three weeks old and have not yet been peer-reviewed in the traditional sense. Both of those things are true simultaneously.