OpenAI's Astra Proves 10 Math Problems Stuck for a Decade

OpenAI published machine-verified solutions to ten open problems in mathematics, each unsolved for at least a decade, generated by an internal research model the company is calling Astra. The centerpiece: an explicit construction of a non-sofic group — an existence problem open since 1999.

Key facts

• Ten proofs, one model. All ten results come from Astra’s internal research pipeline, described by OpenAI as its next-generation model still under review and not publicly released.
• Headline result: the construction of a non-sofic group answers a question that has stood open for more than 25 years.
• Machine-verified, not just plausible. The proofs ship with formal Lean 4 certificates, meaning a computer checked every inference step — the claims do not rest on handwritten “trust me” mathematics.
• Small compute, big yield. OpenAI reports the entire effort ran on roughly $2,000 of compute at current API rates — a cost figure that reframes the economics of mathematical research.
• Publication: the proofs are being released as a 249-page manuscript collection alongside the Lean 4 proof certificates and the model’s reasoning trace for community verification.
• No public release yet. A full release of Astra is contingent on the completion of the US government security review required before such systems are exposed; OpenAI declined to give a timeline.

The announcement

The blog post — Ten Advances in Mathematics — describes the work as the first time a frontier AI system has produced original, formalized, machine-verifiable mathematics at volume. The team writes that the proofs span group theory, combinatorics, and number theory, and that Astra’s recorded reasoning (“scratch-pad”) is being published alongside the certificates, precisely so external mathematicians can audit not just the theorems but the path to them.

The structural pattern inside the results matters as much as the results themselves: at least two proofs introduce novel intermediate lemmas rather than repackaging known techniques — the signature of a system that discovered something rather than retrieved a memory of it.

Why it matters

• Mathematics becomes verifiable AI output. With Lean 4 certificates, claims can be machine-checked end-to-end, which sidesteps the reproducibility crisis currently hurting AI research claims. Nobody can accuse a certified proof of being embellished.
• The $2,000 question. If frontier-grade mathematical progress costs about what a small consulting invoice does, the bottleneck moves to problem selection, proof checking, and human interpretation — not compute.
• A new standard for frontier safety disclosure. Publishing reasoning traces with results sets an expectation that capability claims come with evidence, which is directly relevant to the security-review debates around agentic systems (see our coverage of the OpenAI breach report and the Kill Switch Act).
• Positioning for Astra. The math milestone is the first public capability demonstration of the model line that OpenAI says must pass federal security review before any public rollout.

What to watch

• Peer review success. The field is treating each of the ten results with careful skepticism; the question is which proofs survive contact with the mathematicians who will reproduce them.
• Security review outcome. Absent of a release date for Astra, watch for signals from the US government security review process.
• Formal methods adoption. Whether Lean (and the Lean-verify feedback loop) becomes the default standard for other labs’ AI math claims.

Official source

• OpenAI research blog: Ten Advances in Mathematics
• Sam Altman’s X account: @sama
• Proof certificates: Lean 4 + notebooks repo via OpenAI’s publication

Updated August 8, 2026 — security review and peer reproduction ongoing.