The internal version of the new Astra family of AI models has achieved an impressive result, solving or significantly advancing 10 long-standing open problems in mathematics and theoretical computer science. Public access to the model is not yet available, but the company has already published detailed materials, including a 249-page paper, reasoning breakdowns, and a repository with formalized proofs in Lean 4.

The essence of the claimed breakthrough

According to my analysis of the published data, OpenAI claims that Astra has tackled problems in high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. Notably, the token costs for finding solutions amounted to only about $2,000 at Sol API rates—a sum that underscores a radical reduction in the cost of scientific research.

The key point is honest attribution. OpenAI explicitly stated that attributing authorship of a proof fully generated by AI to a human would be a distortion of the system's contribution. This is an important precedent for the discussion about the role of AI in science.

Significance for quantum topics and cryptography

Two results are of particular interest. First, Astra proved exponential parallel repetition for arbitrary finite two-player quantum games—this could have implications for quantum verification and interactive proofs. Second, the model obtained a stronger result on the complexity of the closest vector problem, which is critical for lattice cryptography—one of the pillars of post-quantum protection. However, I emphasize: this does not mean a threat to already deployed post-quantum systems.

Among other achievements

  • Construction of a non-sofic group—the main result of the publication;
  • Improved estimates for sphere packing in high dimensions;
  • Stronger bounds for binary and spherical codes;
  • New lower bounds in arithmetic circuit complexity;
  • A result on Ehrhart's volume conjecture;
  • A counterexample to Connes's rigidity conjecture;
  • New results on multicolor Ramsey numbers and counterexamples to Erdős's conjectures.

The need for verification

Formalization in Lean is a powerful tool that reduces the risk of logical errors, but it does not replace scientific expertise. Mathematicians will need to assess how well the formal statements correspond to the original problems and how they fit into the existing literature. This is especially important in light of the Leiden Declaration on AI and mathematics, which calls for preserving peer review and transparent attribution.

It is telling that OpenAI did not disclose the number of failed runs or name a date for the public release of Astra. The reproducibility of the claimed process cannot yet be independently verified. This breakthrough recalls recent cases where AI has already refuted 80-year-old hypotheses, but here the scale and depth of the results reach a fundamentally new level.

My conclusion: we are witnessing a turning point. If the results are confirmed, this will accelerate the transition from AI as a tool to AI as an independent researcher. However, until the model is available for external testing, the claims should be treated with cautious optimism.