Ten problems. Each open for at least a decade. Roughly $2,000 of compute to crack all of them.
That was OpenAI’s claim on August 1. The claim is not the interesting part.
The interesting part is that the company published machine-checkable proofs on GitHub, so nobody has to take its word for anything.
What Was Announced
OpenAI said an internal version of Astra, described as its next major model, produced ten new results in mathematics and theoretical computer science.
Alongside the announcement came a 249-page manuscript, a separate account of how the arguments were assembled, and Lean 4 proof certificates for every result. The repository is public under an Apache 2.0 license.
The headline result is the first explicit construction of a non-sofic group. That question has been open since Mikhail Gromov introduced the concept of soficity in 1999.
The other results span von Neumann algebras, high-dimensional geometry, quantum complexity, lattice cryptography and extremal combinatorics.
Why the Lean Files Matter More Than the Results
Lean is a proof assistant. You write a mathematical argument in a formal language and a compiler checks every logical step against a small trusted kernel.
The output is binary. It compiles or it does not. There is no partial credit and no persuasive writing.
That closes the failure mode that has haunted every previous AI mathematics announcement. Models produce arguments that look correct and quietly skip a step, and catching it takes an expert reading closely.
Quartz reported that the repository’s count of unproven placeholders stands at zero, meaning every step across all ten formalizations is fully machine-verified.
Anyone with a laptop can run the check. No PhD required, no peer review queue.
The Caveat That Lean Does Not Cover
A compiled proof guarantees the argument is internally valid. It does not guarantee the statement being proved is the statement mathematicians care about.
Formalizing a problem into Lean involves translation choices. If the formal statement drifts from the informal one, the compiler will happily verify something slightly beside the point.
That check is still human work, and it is underway. None of the ten results has been through a refereed journal process.
So the honest framing is that this clears the verification bar dramatically higher than before, without clearing every bar.
The Quote That Keeps Getting Misattributed
Fields Medalist Timothy Gowers said he would recommend a proof for publication in Annals of Mathematics without hesitation. That line is running in coverage of this announcement.
It was not said about these ten proofs. Gowers said it about a separate result from May 2026, when the same model family disproved the Erdős unit distance conjecture, an 80-year-old problem in discrete geometry.
Several outlets have collapsed the two events. The distinction matters because it is the difference between a mathematician endorsing published work and a mathematician endorsing something he has not finished reading.
Gowers and others have reacted positively to the new results while explicitly noting the work is still being digested.
Thomas Bloom, the Manchester mathematician who curates the Erdős problems catalogue, called the ten results significant, and more so than the earlier unit distance work.
Bloom is a useful voice here for a specific reason. He publicly dismantled an inflated OpenAI claim in October 2025, so he is not a friendly reviewer by default.
| Claim | Status |
| Ten results published with Lean 4 certificates | Verifiable now on public GitHub |
| Every proof step machine-checked | Reported placeholder count of zero |
| Roughly $2,000 compute cost | Vendor-stated, successes only |
| Formal statements match intended problems | Still under human review |
| Peer-reviewed publication | None yet |
| Astra model availability | Unreleased, no date given |
About That $2,000
The figure is the token cost of generating the ten published solutions at OpenAI’s stated API rates.
It excludes every problem the model attempted and failed. That is not a total research cost, it is the cost of the wins measured after the fact.
The number is still striking. But a lab quoting the price of its successes is quoting a floor, not a budget.
OpenAI researcher Sébastien Bubeck, who leads the company’s mathematics research, described the ten as illustrative rather than exhaustive of what the model has produced.
What Nobody Has Confirmed
Astra is unreleased. There is no launch date, no pricing, and no public access.
Some observers have speculated it is the GPT-6 series. OpenAI has described it only as its next major model family, built to coordinate multiple agents on long-running problems.
Treat every capability inference from these proofs as an inference. This is a vendor announcement about a model nobody outside the company can test.
Separately, OpenAI said it is giving 100,000 academic researchers free access to its frontier models through 2027, which deepens the lab’s ties to the research community and concentrates that work on its own infrastructure.
Why This Belongs in a Crypto and Tech Publication
Formal verification is not a mathematics-only tool. It is the same discipline used to prove smart contracts behave as specified, and to prove cryptographic constructions hold.
If models can generate machine-checked proofs at this cost, the economics of auditing formally specified code change with it. Lattice cryptography was one of the six fields covered here, and it underpins post-quantum security work.
Verifiability as a design principle is already familiar territory in this industry, as our explainer on zero-knowledge privacy lays out. The pitch is identical. Prove it mathematically rather than asking for trust.
That is also the argument running through the push for accountable AI as regulators close in on black-box models.
And it fits the broader question of what autonomous systems can be trusted to do unsupervised, which Optimisus examined in why AI agents may become crypto’s next major user base.
The Honest Summary
An AI lab claimed a research breakthrough and then handed over the tools to falsify it. That has not happened often.
The results may hold up completely. The formalizations may need adjusting. Either way, the community can find out in days rather than months.
That process, not the ten proofs, is the actual development worth tracking.
Sources
- The Next Web, OpenAI says its next model Astra has solved ten open problems in mathematics — https://thenextweb.com/news/openai-astra-model-ten-math-proofs-non-sofic-groups
- Quartz, OpenAI Astra model solves ten open math problems, August 3, 2026 — https://qz.com/openai-astra-model-math-problems-lean-proofs-080326
- Tech Times, OpenAI’s Astra solves ten decade-old math problems with machine-checkable Lean proofs, August 2, 2026 — https://www.techtimes.com/articles/322710/20260802/openais-astra-solves-ten-decade-old-math-problems-machine-checkable-lean-proofs.htm
- DataCamp, What the Astra results do and do not establish — https://www.datacamp.com/blog/open-ai-model-astra-solved-ten-open-math-problems
- AI Weekly, OpenAI releases ten Astra math proofs with Lean certificates — https://aiweekly.co/alerts/openai-releases-ten-astra-math-proofs-with-lean-certificates
Optimisus covers crypto and technology news for readers who want the detail behind the headline.
Ava Patel
Ava Patel's expertise lies in creating engaging and informative content about the latest trends and opportunities in the crypto space. Her writing is known for its clarity, accuracy, and ability to convey complex concepts in a way that is accessible to both novice and advanced readers.
