On August 1, 2026, OpenAI Astra — an unreleased internal version of the company's next major model — produced ten results across mathematics and theoretical computer science, each resolving or substantially advancing a problem that had been open for a decade or more. The headline result is the first explicit construction of a non-sofic group, settling a question open since Mikhail Gromov introduced soficity in 1999. Total compute cost: roughly $2,000.
Every proof is formalized in Lean 4 and published on GitHub under Apache 2.0. That last sentence is the actual news, and it is the one getting the least attention.
Because when a proof ships as a machine-checkable Lean certificate, you do not have to trust OpenAI, believe a press release, or wait for a referee. You clone the repository, run the checker, and the claim is either true or it is not. This article covers what Astra proved, why formal verification changes the epistemics entirely, what the $2,000 figure does and does not include, and what it means that the model itself remains locked up pending a US government security review.
Key Takeaways
- OpenAI announced on August 1, 2026 that an internal Astra model produced ten results on long-open mathematical problems, each unsolved for 10+ years.
- The flagship result is an explicit construction of a non-sofic group, open since 1999.
- Every proof is formalized in Lean 4 and published on GitHub under Apache 2.0 — independently verifiable without trusting OpenAI.
- Stated compute cost was about $2,000, which covers inference on these problems and not the training of the model that produced them.
- Astra is still private, has no release date, and must clear a US government security review before any public rollout.
What did OpenAI Astra actually prove?
Astra produced ten results spanning group theory, von Neumann algebras, high-dimensional geometry, quantum complexity, lattice cryptography, and extremal combinatorics. OpenAI published them alongside a 249-page technical manuscript and a separate 62-page account of the process.
The result drawing the most attention among mathematicians is the non-sofic group construction. A brief explanation of why it matters, without the machinery:
Sofic groups are, loosely, groups that can be approximated by finite permutations — every finite piece of the group's structure can be mimicked closely by shuffling a finite set of objects. Gromov introduced the concept in 1999, and a natural question followed immediately: does every group have this property? For 27 years nobody could produce a group that fails it, and nobody could prove that none exists. A great many theorems were proved for sofic groups precisely because the class might be everything.
Astra produced an explicit construction of a group that is not sofic. That does not merely answer a yes/no question — it means the class of sofic groups is a genuine restriction, and every theorem carrying "sofic" as a hypothesis now has a known boundary.
The remaining nine results are spread thinly across fields, which is itself notable. The Next Web's coverage emphasized the non-sofic construction; the breadth is arguably the stronger signal, because a model that only performed well in combinatorics would suggest pattern-matching on a well-trodden literature rather than general capability.
Why Lean 4 formalization is the real story
Because it removes trust from the equation. A Lean 4 proof is checked by a program, not by a person, and the checker does not care who or what wrote the proof. Either every inference step type-checks against the axioms or it does not.
This is a categorical change from how the previous round of AI mathematics results were received. Earlier in 2026, an internal OpenAI reasoning model disproved Erdős's planar unit distance conjecture — a genuinely significant result that nonetheless left mathematicians arguing about verification, understanding, and what it means for a proof to arrive by unexplained means. We covered that debate in AI solved Erdős problems: what actually happened.
Formalization short-circuits the first half of that argument. Compare the two verification pipelines:
| Informal proof (natural language) | Lean 4 certificate | |
|---|---|---|
| Who verifies | Human referees, over months | A type checker, in minutes |
| What you must trust | The referee's attention and expertise | The Lean kernel and your axiom set |
| Failure mode | Subtle gap survives review | Does not compile |
| Reproducible by anyone | No — requires domain expertise | Yes — requires a laptop |
| Vendor claims matter | Yes | No |
The practical consequence: OpenAI could not have faked these results if it wanted to. Publishing under Apache 2.0 on GitHub means any researcher, journalist, or skeptic can run the check themselves. That is a much stronger position than "trust our benchmark," and it sets a standard other labs will now be measured against.
What formalization does not settle is the second half of the argument — whether producing a correct proof by unexplained means constitutes mathematical understanding. A Lean certificate proves the theorem. It does not explain why the theorem is true in a way a mathematician can absorb, generalize, and build on. That objection survives intact, and it is a serious one.
What does "$2,000 in compute" actually mean?
It means the inference cost of generating these ten results, not the cost of the capability that produced them. The distinction is large enough that quoting the figure without it is misleading.
Forbes and others led with the $2,000 number, and it is genuinely striking — a decade-old open problem in group theory resolved for less than a laptop costs. But three caveats belong next to it:
- Training is excluded. Astra is a frontier model. The cost of training it is somewhere between eight and ten figures, and the $2,000 buys access to that investment, not a replacement for it.
- Failed attempts are unclear. OpenAI reported ten results. Whether the $2,000 covers only successful runs, or all exploration including dead ends, is not specified in a way that can be checked.
- Problem selection is invisible. Ten results were published. How many problems were attempted is not disclosed, and the ratio matters enormously for interpreting the capability.
None of this makes the achievement smaller. It makes the marginal cost of mathematical research, given a frontier model, look startlingly low — which is the real economic claim and a genuinely important one. It just is not the same claim as "mathematics now costs $2,000 a problem."
Why is Astra still locked up?
Astra has no release date and must pass a US government security review before any public rollout. The proof files are public; the model that produced them is not.
This is the detail with the longest shadow. OpenAI is publishing frontier research results from a system it has decided the public should not yet have — and the gating factor is a government process, not an internal safety timeline. That is a meaningful shift in how frontier models reach users, and it lands in the same year the industry has been arguing about who controls access to advanced capability.
One of the ten results touches lattice cryptography, which is the mathematical foundation of most post-quantum encryption standards. A model demonstrably capable of producing novel results in that area is not a neutral artifact from a national security perspective, whatever the specific result turns out to say. That alone explains a review process better than any general policy argument.
For developers, the practical takeaway is narrow: nothing here is available to you. The current shipping frontier is what you can actually build on — see our breakdowns of GPT-5.6 Sol, Terra, and Luna and Claude Opus 5 for that. Astra is a preview of the ceiling, not a tool.
What this changes for research
The shift is from AI as a search tool to AI as a producer of verifiable artifacts. Those are different roles with different consequences for how research gets done.
Three things follow, in the order they will probably arrive:
- Formalization becomes the default for AI-assisted results. Once one lab ships Lean certificates, "we have a proof but it is not formalized" becomes a weaker claim than it was in July. Expect competitors to match this within months.
- The referee bottleneck moves. If machines check correctness, human review shifts toward significance, exposition, and whether the result is interesting — which is where mathematicians would rather spend their time anyway.
- Problem selection becomes the scarce skill. If generating a proof costs $2,000 and verifying it is automatic, the valuable human contribution is knowing which questions are worth asking.
That third point is the one I would bet on. The bottleneck in mathematics was never checking; it was taste.
Frequently asked questions
What is OpenAI Astra? Astra is an unreleased internal model that OpenAI describes as its next major system. On August 1, 2026 the company announced that Astra produced ten results on long-open problems in mathematics and theoretical computer science. The model is not publicly available and has no release date.
What is a non-sofic group and why does it matter? Sofic groups can be approximated by finite permutations, a property introduced by Mikhail Gromov in 1999. Nobody had produced a group failing that property, so it was unknown whether the class was a genuine restriction. Astra's explicit construction of a non-sofic group settles that, and it bounds every theorem that assumed soficity as a hypothesis.
Are the Astra proofs actually verified? Yes, in the strongest available sense. Every proof is formalized in Lean 4 and published on GitHub under the Apache 2.0 licence, which means anyone can run the proof checker independently. Verification does not depend on trusting OpenAI's claims.
Did the Astra proofs really cost only $2,000? That figure covers the compute used to generate these ten results. It excludes the cost of training Astra itself, which is orders of magnitude higher, and OpenAI has not disclosed how many problems were attempted unsuccessfully. The $2,000 is a marginal cost, not a total one.
When will OpenAI Astra be released? There is no announced release date. OpenAI has said the model must pass a US government security review before any public rollout, and one of the published results touches lattice cryptography, which underpins post-quantum encryption standards.
Is this different from the Erdős conjecture result earlier in 2026? Yes, in the crucial respect. The earlier result — disproving Erdős's planar unit distance conjecture — was an informal proof that human mathematicians had to verify, and the community argued about it. Astra's results ship as machine-checkable Lean 4 certificates, so correctness is settled before the argument starts.
The verdict
Ten open problems is a remarkable haul and the non-sofic construction is a genuine landmark. But the durable contribution here is procedural, not mathematical: OpenAI shipped proofs that do not require anyone to trust OpenAI.
That is the standard the field should adopt and hold. Benchmark scores are self-reported; Lean certificates are not. Any lab claiming a mathematical result from here on can be asked a simple question — where is the formalization? — and the absence of an answer will mean something specific.
Mathematics just got a verification layer that does not care who wrote the proof. That matters more than the ten theorems.
