On 1 August 2026, OpenAI previewed its next major model family, Astra, by publishing ten solutions to mathematical problems that had been open for at least a decade, one of them for close to 30 years. The unusual part is not the claim but the packaging: alongside a 249-page manuscript, OpenAI released machine-checkable Lean 4 proof certificates to a public repository, so anyone with a Lean toolchain can verify the mathematics without access to the model. The stated inference cost for finding all ten solutions was roughly $2,000 at the company's Sol API rates.
TL;DR
- OpenAI announced the OpenAI Astra model on 1 August 2026 via ten new mathematical results rather than a benchmark chart (openai.com).
- Every result ships with a Lean 4 certificate in github.com/openai/ten-proofs (Lean 4.32.0, mathlib, Apache-2.0), so correctness is independently checkable.
- Reported inference spend to find all ten solutions: about $2,000. That figure excludes the wider research programme.
- Astra itself is not released. No public API, no date, and outside researchers cannot run it.
- Astra is slated to be the first model reviewed under a new U.S. classified frontier threshold (CFT) process before public release.
- The transferable lesson for engineering teams: pair generated output with an automated checker, because the checker is what makes the output trustworthy.
What exactly did the OpenAI Astra model do?
Astra generated the mathematical arguments for ten previously open problems. Humans then helped prepare the written manuscripts, using the same model, and the model formalised each argument in Lean 4. OpenAI's framing is deliberate on authorship: the arguments were produced by the system, humans handled write-up and formalisation, and OpenAI takes responsibility for correctness. The company points to the Leiden Declaration on AI and Mathematics and states plainly that claiming human authorship for an AI-generated proof would misrepresent both the system's contribution and what human intellectual work actually involves.
The ten results span several distinct fields rather than clustering in one convenient corner of combinatorics:
| Area | Result |
|---|---|
| High-dimensional sphere packing | New upper bounds on packing density down to the Cohn-Elkies threshold, the first improvement to the general sphere-packing exponent since 1978 |
| Coding theory | Exponentially improved bounds on the maximum size of binary codes at a prescribed minimum distance |
| Group theory | Construction establishing that non-sofic groups exist, a question posed by Gromov around 1999 |
| Operator algebras | Disproof of the Connes rigidity conjecture: some groups are not uniquely determined by their von Neumann algebras |
| Circuit complexity | New arithmetic circuit lower bounds for the permanent, including an arithmetic-formula bound of order n⁴/log n |
| Quantum complexity | An exponential parallel repetition theorem for general two-player quantum games |
| Lattice problems | Polynomial-factor hardness of approximation for the closest vector problem |
| Convex geometry | Ehrhart's volume conjecture, in every dimension |
| Ramsey theory | A superexponential lower bound for multicolour triangle Ramsey numbers, resolving Erdős problem 183 |
| Extremal graph theory | Compactness and degeneracy results, resolving Erdős problems 146 and 180 |
Two of those have direct engineering relevance. Hardness of approximation for the closest vector problem sits underneath lattice-based post-quantum cryptography, and improved binary code bounds touch error-correction theory. Neither result breaks a deployed system; both sharpen the theoretical picture that security proofs lean on.
Why do the Lean 4 certificates matter more than the maths?
Because they change what a capability claim costs to check. A conventional announcement asks you to trust a benchmark score you cannot reproduce, on a model you cannot run. A Lean certificate asks you to run a compiler. The repository was created at 06:10 UTC on 1 August 2026 and pins Lean 4.32.0 with mathlib under an Apache-2.0 licence. If the proofs typecheck, the theorems hold, regardless of what you think of the model that produced them.
This is a visible shift from OpenAI's earlier maths announcement. On 20 May 2026 the company reported an AI-generated disproof of the Erdős unit-distance conjecture, open since 1946. That result depended on expert sign-off, which is slower and harder for outsiders to audit. Moving to machine-checkable artefacts removes the human bottleneck from verification, though not from interpretation: Lean tells you a proof is valid, not whether the statement formalised is the statement mathematicians care about. Reading the formal statements remains necessary work.
Reaction from people who track these problems was positive. Thomas Bloom of the University of Manchester, who maintains erdosproblems.com, described the release as bigger news than the May unit-distance result. OpenAI researchers Noam Brown and Sebastien Bubeck both commented publicly, with Bubeck confirming the non-sofic groups result.
Does $2,000 per breakthrough mean research is now cheap?
Not in the way the headline suggests, but the number still matters. The $2,000 covers inference to find all ten solutions at Sol API rates, not training, manuscript preparation, formalisation, or failed attempts. Treating it as the full cost of ten breakthroughs would be a category error.
What the figure does establish is a new floor for the marginal cost of an attempt. When a serious try at a hard open problem costs a few hundred dollars of compute rather than months of a specialist's time, the economics of which problems get attempted change. Speculative, low-probability attacks on long-standing conjectures become affordable in bulk. That is a different claim from "maths is solved", and it is the more useful one for anyone planning research budgets. It also mirrors a pattern we have covered elsewhere: raw capability gains do not translate straight into productivity, as the gap between AI coding tools and actual developer throughput keeps demonstrating.
What does the government review step change?
Astra is set to be the first model routed through a new U.S. classified frontier threshold framework before public release. An executive order signed on 2 June 2026 gave three agencies 60 days to stand up a benchmarking process, with the deadline landing on 1 August 2026. In the week of 26 July, Sam Altman demonstrated Astra to politicians and regulators in Washington, D.C. The move towards government-gated releases also fits the broader shift in LLM training paradigms from web text to reasoning priors.
The practical consequence is that a frontier launch now has a clearance dependency. Release timing becomes partly a policy question rather than purely an engineering one, and that cuts both ways for the ecosystem. It adds a gate that open-weight releases do not face, which is one reason the contrast with something like Kimi K3's open-weight approach is worth watching. It also means capability previews of this kind, where evidence is published but the model is withheld, may become the normal shape of an announcement.
What should engineering teams take from this?
The reusable idea is verification-first output. Astra's results are credible because a checker validates them mechanically, not because a lab vouched for them. That pattern generalises to work that has nothing to do with pure mathematics: property-based tests, typed interfaces, schema validation, and formal specifications all serve the same function of making generated output auditable rather than merely plausible.
Concretely, if you are putting a model to work on something consequential, the useful question is what artefact proves the work is correct. A Lean proof, a passing test suite, a satisfied contract, or a reproducible build all qualify. Prose explanations do not. This is also where the skills gap around AI agents shows up most sharply: the scarce ability is not prompting, it is designing the check.
Two honest limitations. First, Astra is unavailable, so none of its reasoning behaviour can be independently probed. Second, formal verification is far easier in domains with a mature formal library, which mathlib provides for mathematics and few fields have an equivalent for. The method demonstrated here is not portable to arbitrary problem domains today.
FAQ
Q: Can I use the OpenAI Astra model right now? A: No. Astra is internal and unreleased, with no public API or release date. Only the outputs — proofs, manuscript, and reasoning walkthroughs — are public.
Q: How can the results be verified without access to the model? A: The Lean 4 certificates in github.com/openai/ten-proofs can be compiled with Lean 4.32.0 and mathlib. If they typecheck, the theorems hold — verification depends on the Lean toolchain, not on OpenAI.
Q: Was the $2,000 the total cost of the research? A: No. It covers inference at Sol API rates to find all ten solutions. Training costs, manuscript preparation, formalisation effort, and failed attempts are excluded.
Q: Do these results break post-quantum cryptography? A: No. The closest vector problem result concerns hardness of approximation and strengthens lattice-based scheme theory. It is not an attack on deployed systems.
Q: Did humans contribute to the proofs? A: The mathematical arguments were generated by the system. Humans helped prepare manuscripts and Lean formalisation; OpenAI takes responsibility for correctness but declines to claim human authorship of the arguments.
Q: How is this different from the May 2026 Erdős result? A: The May disproof relied on expert review. This release provides machine-checkable Lean certificates, so correctness is confirmed automatically.
Discussion
0 comments