OpenAI reveals next model Astra by solving 10 open math problems
The proofs are public and Lean-verified — the model isn't. Astra faces the first US federal safety review, and its $2K cost counts only the wins.
Copy markdown
Ten decade-old problems, ten machine-checked proofs
Astra produced solutions to 10 problems open for 10+ years — the first non-sofic group construction in 27 years, Connes's 1980 rigidity conjecture, Ehrhart's volume conjecture and three Erdős problems among them. All ten ship as Lean 4 certificates with a 'sorry' count of zero, so every step is formally verified.
It's a multi-agent model, not a chatbot
OpenAI calls Astra its "next major model family": one that coordinates multiple agents on a single problem for hours or days rather than answering in one pass. It hasn't decided whether it ships as GPT-6 or a GPT-5 variant.
You can't use it — Washington decides when you can
Astra is unreleased with no date, and it's the first model that must clear the new US federal AI-safety review before public launch. The real builder takeaway: the next capability tier provably exists, but access is now federally gated.
That $2,000 counts only the wins
The viral "~$2,000 for 10 problems" figure is hypothetical at Sol API rates and covers successes only — OpenAI's Noam Brown said they "did try other major problems without success." Worst-case estimates land nearer ~$20K per major proof.
Machine-checked isn't peer-reviewed
The Lean certificates prove the arguments are formally valid, but no mathematician has vetted the results — specialist digestion "takes months." Treat this as verified-correct, not yet community-accepted.
Build this weekend: mine the proof repo
The 249-page manuscript and every Lean 4 certificate are on GitHub under Apache 2.0. If you work in formal verification or AI-for-math, you can rerun the proofs through Lean, study the model's search, or train on the traces — no Astra access required.