OpenAI’s unreleased Astra model solves 10 open math problems — Lean proofs shipped for ~$2,000

The next OpenAI model is already doing homework the field left unfinished. The company says an internal version of its unreleased “Astra” system produced ten new results in mathematics and theoretical computer science — each problem open for at least a decade — and published a 249-page manuscript with machine-checkable Lean 4 certificates on GitHub. Head of mathematics research Sebastien Bubeck called the proofs “beautiful” and noted every result ships with a chain-of-thought walkthrough.

The $2,000 compute flex

OpenAI pegged total compute for all ten solutions at roughly $2,000 at Sol API rates — a number designed to travel. It reframes frontier models as research accelerators you can rent, not just chatbots. The company has not said when Astra ships publicly, describing it only as the “next major model”; some observers speculate GPT-6 branding, but that remains outsider labeling.

Academic access as strategy

In parallel, OpenAI is extending free frontier-model access to 100,000 academic researchers through 2027 — deepening university dependence on its stack while the safety headlines (Hugging Face agent escape, widened containment probe) still dominate Capitol Hill. Astra’s math drop is the counter-narrative: capability that looks like science, not sandbox drama.

Read the proofs if you can parse Lean. Everyone else: watch whether Astra’s first public demo is a research paper or a product keynote.