OpenAI says unreleased Astra model solved 10 open math problems

Why it matters

Astra gives OpenAI its clearest evidence that frontier models may generate auditable research results, opening a market for AI systems sold as discovery infrastructure.

OpenAI said on August 1st that an internal version of Astra, its next major model, produced new results for 10 longstanding problems in mathematics and theoretical computer science, a claim that pushes the frontier AI contest beyond benchmark scores and into work that can be inspected by research mathematicians.

Noam Brown (@polynoamial), an OpenAI research scientist, wrote on X that Astra represents a "major step for scientific reasoning." Brown works on reasoning, reinforcement learning, self-play and multi-agent systems. Before joining OpenAI, he worked at Meta's FAIR lab on CICERO, an AI system built to play the strategy game Diplomacy, and at Carnegie Mellon University on the Libratus and Pluribus poker systems.

The evidence package is larger than the benchmark tables that usually accompany a frontier model announcement. OpenAI published a 249-page collection of manuscripts, reasoning walkthroughs and a public GitHub repository of Lean certificates. Lean is a proof assistant that can mechanically check whether a formal argument follows from its stated assumptions.

According to OpenAI's research announcement, Astra generated the mathematical arguments. Humans then used the same model to prepare them as manuscripts, after which Astra formalized each argument in Lean. OpenAI said it takes responsibility for the correctness of the work while crediting the system with generating the underlying arguments.

The 10 claimed advances cover high-dimensional sphere packing, binary and spherical codes, non-sofic groups, Connes's rigidity conjecture, arithmetic circuit complexity, quantum parallel repetition, the closest vector problem, Ehrhart's volume conjecture, multicolor Ramsey numbers and two conjectures in extremal graph theory.

Several carry consequences outside pure mathematics. The closest vector problem is central to lattice theory and post-quantum cryptography. Arithmetic circuit lower bounds concern fundamental limits on computation. Quantum parallel repetition addresses how success probabilities behave when quantum games are repeated, a question tied to quantum complexity theory.

The most striking claims include an explicit construction of a non-sofic group, which would resolve whether every countable group can be approximated by finite permutations, and a disproof of Connes's rigidity conjecture. The sphere-packing result claims the first improvement since 1978 to the general high-dimensional packing exponent.

OpenAI said the tokens used to find the 10 solutions would cost roughly $2,000 at GPT-5.6 Sol API rates. That figure translates inference tokens into a current product price. It does not represent Astra's training cost or the human work required to select problems, prepare manuscripts and check the output.

The model itself remains internal. OpenAI identified Astra as its next major model but did not give a release date, pricing, access terms or a model card. The announcement therefore functions as a capability preview backed by artifacts that outside specialists can audit, rather than a model launch.

The breadth of the claims separates Astra's showing from the math demonstrations frontier labs were publishing a year earlier. Those releases often centered on Olympiad questions or curated benchmarks with known answers. Research mathematics requires a system to identify a viable route through technical literature, produce an original argument and survive detailed review.

Google DeepMind has been pursuing the same transition. Its Gemini Deep Think research program uses a math agent called Aletheia to generate and revise solutions under expert direction, following earlier gold-medal-level results on student competitions. OpenAI's Astra announcement raises the competitive bar by presenting a batch of proposed original results across several fields at once.

The Lean certificates provide a stronger checking mechanism than natural-language proofs alone, but they do not settle every question surrounding the release. Researchers still need to examine whether each formal statement captures the intended open problem, whether the assumptions are appropriate and how important the resulting advance is within its field. A proof can be mechanically valid while its mathematical significance still requires expert judgment.

OpenAI has been building toward this claim. In May, the lab disclosed that an unreleased model had generated a disproof of the Erdos unit-distance conjecture, an 80-year-old geometry problem. Astra turns that earlier result into a wider product thesis: the next frontier model is being built as research infrastructure, with scientific discovery positioned as a core capability rather than a specialist demonstration.