An unreleased OpenAI model helped solve ten longstanding maths problems

Subject: AI-assisted mathematics is moving from conjectures to publishable proofs.

Preview: OpenAI says its internal Astra model contributed to ten mathematical advances, with human researchers writing and formalising the resulting proofs.

A strong model can accelerate discovery, but scientific credit still depends on verification, formalisation and expert judgement.

OpenAI reported ten results on problems that had seen no progress for at least a decade, including work on non-sofic groups, Connes rigidity and three Erdős problems. The internal Astra model supported the work; researchers wrote the papers and several proofs were formalised in Lean.

This points to a more useful role for advanced AI in science: not replacing researchers, but expanding the number of promising routes they can explore and verify.

A model producing a plausible line of reasoning is not the same as a solved theorem. The durable result comes from human scrutiny, formal proof and publication, not from the model output alone.

For research workflows, separate idea generation from validation. Record prompts and intermediate reasoning, require expert review, and use formal tools where the domain allows it.

The breakthrough is the collaboration pattern: AI proposes and explores; researchers establish what is actually true.