OpenAI Introduces Astra With Ten Solved Math Problems

OpenAI's next major model has a name, and it arrived without a launch event. On August 1, the company announced Astra by publishing proofs for ten open mathematics problems, questions that had resisted mathematicians for at least a decade, in some cases far longer. No release date, no product page, no demo video. Just the proofs.

What exactly was solved?

The ten results span high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. Two stand out. Astra proved that non-sofic groups exist, settling a central open question in group theory, and it refuted the Connes Rigidity Conjecture in the theory of von Neumann algebras. The list also includes new upper bounds on sphere packing density and exponentially improved bounds on the maximum size of binary codes.

The verification method matters as much as the results. Every argument was formalized in Lean, a proof assistant that checks each logical step by machine, and OpenAI published the certificates alongside chain-of-thought walkthroughs in a public GitHub repository. Researcher Sébastien Bubeck framed the announcement bluntly: "yes, nonsofic groups exist: this statement is one of many new beautiful results proved by Astra, our next major model."

What we know about the model itself

Not much, and OpenAI seems comfortable with that. The proofs came from an internal version of Astra; the model has not shipped, and it may ultimately launch under a different name. What has been confirmed: Astra is built for long-horizon tasks that run for hours or days, it will be the first model submitted to the new US government review framework, and Sam Altman has already demoed it for policymakers in Washington. Noam Brown added two revealing details: the compute spent per problem was modest, roughly $2,000 in total at current API prices, and the team tried bigger targets, including Millennium Prize problems, without success.

Reading the announcement as a business signal

Formal mathematics is about as far from a customer support queue as it gets, so why should anyone outside academia care? Because machine-checked proofs are the cleanest possible evidence of a capability that transfers: sustaining a long, error-intolerant chain of reasoning to a verifiable conclusion. That is the same skill profile behind large codebase migrations, multi-step compliance audits, and autonomous workflows that run overnight. When this class of capability reaches public APIs at commodity prices, tasks currently labeled too risky for AI will get re-priced. Companies that have already broken their processes into verifiable steps will move first.

One caution belongs in the record: this is an announcement strategy, not peer review. The Lean certificates make the proofs trustworthy, but the mathematical community will need time to absorb the results, and Astra's performance on messier real-world work remains unproven.

Sources: OpenAI, The Decoder, GitHub: ten-proofs