OpenAI reveals Astra with ten Lean-checked proofs of open problems
The proofs ship as an Apache-2.0 Lean repo you can check; the model stays locked, ~$2K in tokens did it, and a new federal gate awaits.
Copy markdown
Ten open problems, each proof formalized in Lean
Astra cracked ten previously open results across math, quantum complexity, and theoretical CS — including a disproof of Connes' rigidity conjecture and the first sphere-packing bound improvement since 1978. Every proof ships as an Apache-2.0 Lean 4.32 certificate at github.com/openai/ten-proofs, so you can machine-check the claims yourself.
The catch: you can check the proofs, not the model
Astra is OpenAI's unreleased next model, and it stays strictly internal — no API, no waitlist, no release date. Only the certificates are public; the model, methods, and training data can't be independently tested. OpenAI hasn't even decided whether it ships as GPT-6 or a GPT-5.7-style variant.
~$2,000 in tokens solved decade-open problems
By OpenAI's own estimate, generating all ten solutions would run about $2,000 at Sol's API rates — a concrete marker of how cheap frontier-grade reasoning has gotten. Whatever Astra ships as, that's the compute price band behind results humans spent years chasing.
Astra reportedly faces a new federal pre-release gate
Reports say Astra will be the first model routed through the Trump administration's near-final frontier-model framework, which requires new frontier models to be submitted to the government before public release. For builders, that's a fresh checkpoint that could delay when the next big model reaches your API keys.
Elsewhere: California's AI Transparency Act goes live
As of today, gen-AI providers with 1M+ California users must embed C2PA provenance in every AI image, video, and audio and run a free public-plus-API detection tool — $5,000 per day per violation. If you ship generative media at scale, content provenance is now a legal default, not a nice-to-have.