OpenAI’s Astra Solves 10 Long-Standing Math Problems Ahead of Public Release

OpenAI has announced that OpenAI Astra, its next-generation AI model family currently under development, solved ten long-standing open problems in mathematics and theoretical computer science during internal testing. According to the company, the model generated solutions at an estimated inference cost of about $2,000 using Sol API pricing before researchers converted the results into formal academic manuscripts.

OpenAI also published machine-verifiable proof certificates and reasoning records, allowing researchers to examine the model’s findings.

OpenAI Astra Tackles Decades-Old Problems

According to OpenAI, OpenAI Astra solved problems that had remained unresolved for at least a decade across several mathematical and computer science disciplines. These included high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics.

After generating the solutions, researchers used the same model to help prepare academic manuscripts. Each proof was then formalised in Lean, a language used for machine-verified mathematical proofs, before OpenAI released the certificates publicly.

Major Mathematical Results

OpenAI said the model produced several significant results, including:

• New upper bounds on sphere-packing density reaching the Cohn-Elkies threshold.
• A construction demonstrating the existence of non-sofic groups, addressing a question open since 1999.
• A disproof of Connes’s rigidity conjecture involving von Neumann algebras.
• Stronger mathematical lower bounds for the computational complexity of calculating the permanent, an important problem in arithmetic circuit complexity.
• A new theorem showing how the difficulty of repeated two-player quantum games increases as the games are played multiple times.
• Polynomial-factor hardness of approximation for the closest vector problem, which has implications for post-quantum cryptography.
• A proof of Ehrhart’s volume conjecture in every dimension.
• A superexponential lower bound for multicolour triangle Ramsey numbers, resolving Erdős problem 183.
• Results addressing compactness and degeneracy conjectures in extremal graph theory, resolving Erdős problems 146 and 180.

Commenting on the announcement, University of Manchester mathematician Thomas Bloom described the work as “big news” and a “significant step.”

Astra Has Not Been Released

Despite the announcement, OpenAI Astra is not yet available to the public.

OpenAI CEO Sam Altman reportedly demonstrated the model to U.S. senators and regulators in late July. According to The Information, the company has not decided whether Astra will launch as GPT-6 or become part of the GPT-5 model family, and no public release date has been announced.

The model could also become one of the first AI systems evaluated under the proposed U.S. pre-release AI assessment framework.

OpenAI Clarifies AI’s Role

OpenAI acknowledged that OpenAI Astra did not solve the Millennium Prize Problems, a collection of seven major unsolved mathematical challenges established by the Clay Mathematics Institute.

The company also stated that claiming human authorship for proofs generated entirely by AI would misrepresent both the system’s contribution and the nature of genuine human intellectual work, highlighting the importance of transparency in AI-assisted research.

Key Highlights

Features	Details
Model	OpenAI Astra
Status	Internal, unreleased
Problems Solved	10 long-standing mathematics and computer science problems
Verification	Proofs formalised using Lean
Estimated Inference Cost	Approximately $2,000 (Sol API rates)
Release Date	Not announced
Possible Launch	GPT-6 or GPT-5 family (not yet confirmed)

Conclusion

OpenAI’s announcement suggests that OpenAI Astra has demonstrated advanced mathematical reasoning by solving ten long-standing problems across multiple research fields. While the results have been publicly documented with machine-verifiable proofs, the model remains unreleased, and its future product positioning has not been confirmed. If independently validated by the wider research community, the work could represent an important milestone in applying AI to complex scientific and mathematical discovery.