OpenAI on August 1 revealed a new frontier model family, Astra, by publishing ten solutions to long-standing open problems in mathematics and theoretical computer science, together with Lean certificates and reasoning walkthroughs. The lab said Astra generated the arguments internally at a total inference cost of roughly $2,000 at Sol API rates.
An Announcement Smuggled Into A Math Paper
The release, titled "Ten advances in mathematics and theoretical computer science," reads less like a product launch and more like a research drop. OpenAI shared manuscripts prepared by human collaborators, formal Lean certificates published to GitHub, and per-problem narrations of Astra's thinking. Fields Medalist Timothy Gowers, cited by early observers, said he would recommend one of the model's proofs to a top journal.
The Ten Results
Astra's results span geometry, coding theory, algebra and post-quantum cryptography. Highlights include new upper bounds for high-dimensional sphere packing that hit the Cohn–Elkies threshold, exponentially improved bounds on binary and spherical codes, a construction proving the existence of non-sofic groups, a disproof of Connes's rigidity conjecture for von Neumann algebras, a superexponential lower bound for multicolor triangle Ramsey numbers, and polynomial-factor hardness of approximation for the closest vector problem – a lattice question central to post-quantum lattice cryptography.
What Astra Actually Is
OpenAI has not committed to a public release date, framing Astra as its "next major model" built to coordinate multiple agents over long-running research tasks rather than answer chat queries. It ties directly into the lab's broader agentic push, following recent agentic and reasoning experiments from OpenAI. The company also positioned Astra alongside IBM's verified quantum advantage claim as evidence that frontier systems are now delivering original scientific contributions.
Why It Matters For Physical AI
Astra's design – multi-agent reasoning that can stay on a problem for hours – is the same architecture powering the wave of foundation models being retooled for robotics and physical AI. The math breakthrough is a stress test that OpenAI can now sustain adversarial reasoning long enough to close proofs, a capability that industry rivals such as Anthropic and Google DeepMind have been racing to match.
Reporting based on coverage from OpenAI, Forbes, The Decoder and Gizmodo.