Skip to content
NewsResearch

OpenAI reveals Astra through ten machine-verified open-problem proofs, headlined by first non-sofic group construction in 27 years

· by Pondero Newsdesk

The short version

OpenAI's Astra produced machine-verified Lean 4 proofs for ten open mathematics and computer science problems on August 1, including the first explicit non-sofic group construction since Gromov introduced soficity in 1999.

OpenAI reveals Astra through ten machine-verified open-problem proofs, headlined by first non-sofic group construction in 27 years

An internal version of Astra, what OpenAI describes as its next major model family, produced machine-verified proofs for ten mathematics and theoretical computer science problems that had resisted progress for at least a decade. The headline result is the first explicit construction of a non-sofic group, closing a question in group theory that stood open since Mikhail Gromov introduced soficity in 1999. OpenAI published the results on August 1, alongside a 249-page manuscript and a public GitHub repository containing machine-checkable Lean 4 certificates for all ten proofs.

The release doubles as OpenAI's formal announcement of Astra. No public release date has been confirmed, and the model will be the first to pass through the TRAINS (Testing Risks of AI for National Security) pre-release government review before launch.

What Astra solved

The ten problems span six fields: group theory, operator algebras, high-dimensional geometry, quantum complexity, extremal combinatorics, and circuit complexity, per The Decoder's coverage.

The non-sofic group result drew the most attention. Soficity is a property describing whether a group can be approximated by finite symmetric groups. Gromov introduced it around 1999 and conjectured that all groups might be sofic. Astra produced the first explicit example proving they are not. For 27 years, no human mathematician had managed it.

The remaining nine: a disproof of the Connes rigidity conjecture in operator algebras (von Neumann algebras); a proof of Ehrhart's volume conjecture in high-dimensional geometry; three problems from Paul Erdos's famous catalog, including problem 183 on multicolored Ramsey numbers; the first improvement to the general upper bound on sphere-packing density in high dimensions since 1978; a parallel repetition theorem for two-player quantum games; and new lower bounds on circuit complexity for computing the permanent, per The Decoder and NextBigFuture.

Total compute for all ten argument sets came to approximately $2,000 at GPT-5.6 Sol API prices. Noam Brown, an OpenAI researcher on the project, noted that figure was a deliberate ceiling: "It's possible to push test-time compute much further," per NextBigFuture. In other words, the $2,000 number reflects what was spent, not a limit on what the approach could produce.

OpenAI also released an LLM-generated PDF in which the model reconstructs how each proof came together using unpublished reasoning traces. The mathematical arguments originated with Astra. OpenAI researchers then worked with the model to formalize them into publishable manuscripts, per Simon Willison's analysis.

Why the Lean 4 certificates change the conversation

Previous AI math demonstrations drew persistent skepticism. Outputs that looked like proofs but could not be independently checked by machine left room for the objection that the model was pattern-matching rather than reasoning. Lean 4 eliminates that objection for these results.

A Lean proof that type-checks is valid by construction. No trust in OpenAI's evaluation claims, benchmarks, or interpretation is required. Any mathematician or independent researcher can pull the certificates from the openai/ten-proofs GitHub repository and verify them directly. That shifts the relevant question: not whether the proofs are correct (they can be checked), but whether the mathematical ideas they contain are new and how they were found.

Sebastien Bubeck, OpenAI's head of mathematics research, called the results "beautiful" on X. Thomas Bloom of erdosproblems.website described them as "big news" and said they are more significant than the unit distance counterexample OpenAI had previously published. Terence Tao, the Fields Medal recipient, responded by articulating a vision of "big mathematics": large-scale decentralized collaborations between humans and machines in which complex research tasks are divided, with humans taking on creative direction and AI handling substantial derivation work, per NextBigFuture.

Astra as a system, not just a model

Astra is designed as a multi-agent system that can coordinate multiple agents working on a single problem across hours or days, a deliberate contrast to current production models that handle discrete, bounded requests. OpenAI has not disclosed the full technical design. What the math results demonstrate is that this type of sustained, long-horizon search can yield original findings in areas where prior AI systems showed no progress.

For teams running AI on research or analysis tasks, the design pattern of having agents work together on a single hard problem over extended time is no longer hypothetical. The sphere-packing and circuit complexity results in particular suggest domains well outside pure math could be next, since both connect to real-world engineering problems in signal processing and computation.

Context: what changed and what has not

Prior AI math systems worked primarily on competition problems (Olympiad-style, where solutions are bounded and verified against known answers) or on problems where human mathematicians had already outlined the path and AI filled in the derivation. The batch published August 1 is structurally different: these are open research problems with no known solution for academic mathematicians to compare against.

Astra did not work in isolation, though. Human researchers directed the formalization process and shaped each result into a publishable manuscript. The precise division between what Astra contributed and what the research team contributed is not spelled out in the announcement. Whether the model's role constitutes mathematical creativity or very sophisticated search over a large learned corpus is a question the academic mathematics community will now spend months examining, starting with the Lean certificates.

What to watch next

Three things will settle how much this matters over the next few weeks.

Whether the Lean 4 certificates survive independent academic scrutiny is the first. Type-checking confirms validity. It does not confirm that the proofs contain genuine insight or that the results are as novel as described. Mathematicians working in group theory, operator algebras, and extremal combinatorics will offer their own readings over the coming weeks.

The TRAINS review is the second gate. Astra is the first model family designated for this pre-release national security evaluation process, established under Executive Order 14409. The review scope and timeline are not public; how long it runs will shape when Astra reaches API access, per The Decoder.

Third, whether OpenAI converts this research preview into a named, versioned product and where Astra fits relative to the GPT series naming remains unannounced. The company has not confirmed whether Astra is GPT-6, a separate product line, or something else.

Sources