OpenAI publishes ten research-mathematics results from its Astra model — with Lean proofs and a reconstruction of how it searched
On 1 August OpenAI released an unusually inspectable package: a 249-page manuscript covering ten results in pure mathematics and theoretical computer science, a second document reconstructing how its internal Astra model found the arguments, and a public repository of machine-checkable Lean proofs. It is the largest single bundle of AI-generated mathematics yet — and the verification story is as important as the results.
The shape of the release is the news. Ten results at once, formalised in Lean so a machine can check every step, plus a companion document that reconstructs the model's search rather than only presenting the polished proof. That combination — results, machine-verification, and a trace of the reasoning — is a deliberate answer to the central objection to AI mathematics: that a proof you cannot inspect is a claim, not a theorem.
The caveat is equally deliberate. Unlike May's Erdős unit-distance disproof, which was announced alongside checks by outside mathematicians, the ten-result bundle does not carry the same broad external-review claim. The Lean formalization means the logic is machine-verified, but independent specialist review of what the results mean and whether they are significant is still pending. Trackers are logging them as provisional, primary-evidence-public, review-pending.
Even provisional, the scale marks a threshold. A frontier lab shipping ten formalised research-mathematics results in a single manuscript reframes what these models are for — not chat, not code, but discovery — and puts the burden on the mathematics community to build the review capacity to keep up with a producer that no longer works at human cadence.
ExplainX — OpenAI Astra's ten math proofs explained: Lean certificates, 2026 → · Kingy.ai — OpenAI Astra's ten math results: evidence and limits →