OpenAI's unreleased Astra solves ten math problems stuck for a decade, at under $2,000 each
OpenAI set an internal version of Astra, its next major model, on ten mathematics and theoretical computer science problems whose main results had stalled for at least a decade, and published Lean 4 formalizations of the solutions in the openai/ten-proofs repository. The price is the signal: under $2,000 per problem at GPT-5.6 Sol token rates, against the $100,000 Anthropic spent finding cryptographic weaknesses days earlier. Simon Willison notes the prompts stay unpublished and the failure count undisclosed, so the real hit rate is unknown — research-grade proof generation is cheap, but its reliability is not yet measurable from outside.
Source: simonwillison.net ↗
They set "an internal version of Astra, our next major model" on finding solutions to ten mathematical problems that "have seen no progress on the main result for at least a decade".
Why this matters
- → Sub-$2k cost for decade-stuck math proofs signals AI research capability reaching new scale
- → Reliability metrics still opaque — solution count and prompt strategy undisclosed
- → Shifts mathematician workflow from solo insight to human-AI collaboration model