In the world of fundamental mathematics, an event has occurred that could change the game. My sources and analysis of internal data indicate that OpenAI has achieved an impressive result: an internal version of their new model, Astra, has obtained solutions or significant progress on ten long-standing open problems in mathematics and theoretical computer science. At the moment, there is no public access to this model, which fuels interest and questions about the claim.
The Essence of the Announced Breakthrough
According to materials published by the company, this involves solving problems in complex fields such as high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, and quantum complexity. Particular attention is drawn to results concerning lattice-based cryptography—one of the pillars of post-quantum security. The company claims that the total computational cost of finding these solutions would have been only about $2,000 at their Sol API rates. This is a strikingly low figure given the scale of the announced discoveries, although it likely does not account for failed runs and subsequent verification work.
Key Achievements and Their Significance
The most significant result, in my assessment, is the construction of a non-sofic group—an object whose existence had remained in question since the late 1990s. Additionally, improved bounds for sphere packing, new lower bounds in circuit complexity, and results on the Ehrhart hypothesis have been claimed. In the context of quantum technologies, the result on the parallel repetition of quantum games and improved bounds for the closest vector problem in lattices are especially important. The latter could have long-term implications for understanding the complexity of breaking post-quantum cryptosystems, although this does not mean an immediate threat to deployed algorithms.
Questions of Attribution and Verification
OpenAI has taken an unprecedented step by publishing a 249-page paper and also formalizing the proofs in the Lean 4 system. This allows for machine verification of the logic. However, the company explicitly states that it does not attribute authorship of the proofs to the people who helped prepare the manuscripts, but rather considers the model itself to be the author. This is a bold position that directly resonates with the discussion of the Leiden Declaration on AI in mathematics. Despite the formalization, the key question remains: how accurately do the formal statements correspond to the original open problems, and how will the mathematical community assess their significance? Without access to the model for independent testing, the reproducibility of this process cannot yet be verified.
My expert commentary: This claim, if confirmed, will become a watershed moment in the history of AI. We are moving from models that assist researchers to systems that independently generate new mathematical knowledge. However, I maintain healthy skepticism: until independent experts verify the formal proofs and cross-check them against the original problems, this is more of a loud demonstration of capabilities than an established scientific fact. The market and the scientific community will be closely watching how events unfold.