OpenAI's Astra Solves 10 Math Problems for $2K

OpenAI’s Astra Model and the 10 Long-Standing Problems It Reportedly Solved

OpenAI says an unreleased internal version of its next model, Astra, produced new results for 10 open problems in mathematics and theoretical computer science, and that each proof was formalized in Lean for machine verification. The claims span geometry, coding theory, group theory, operator algebras, complexity theory, quantum information, lattice cryptography, and extremal combinatorics.

What OpenAI says Astra achieved

According to OpenAI, the internal Astra model found solutions to 10 problems that had remained open for years, then converted each argument into a Lean certificate. The company also says the total token cost to find the solutions was roughly $2,000 at Sol API rates.

• High-dimensional sphere packing, with improved upper bounds on packing density.
• Binary and spherical codes, with exponentially improved bounds.
• Non-sofic groups, with an explicit construction.
• Connes’s rigidity conjecture, which OpenAI says was disproved.
• Arithmetic circuit complexity, with new lower bounds for the permanent.
• Quantum parallel repetition, with an exponential theorem for two-player quantum games.
• Closest vector problem, with hardness-of-approximation results.
• Ehrhart’s volume conjecture, resolved in every dimension.
• Multicolor Ramsey numbers, with a superexponential lower bound.
• Extremal number conjectures in graph theory, resolving two Erdős problems.

Why the Lean verification matters

The Lean component is important because it means the proofs are not just informal claims. Lean is a proof assistant that checks each step of an argument in a machine-readable way, which makes the results easier for mathematicians to inspect and verify independently.

OpenAI says it provided machine-checkable proofs alongside the manuscript collection and reasoning walkthroughs, making the release unusually transparent for an AI-generated mathematics claim.

The most notable problems in the set

Some of the reported results stand out because they address especially famous or long-standing questions.

• Non-sofic groups, a major open question in group theory, were reportedly constructed explicitly.
• Connes’s rigidity conjecture was reportedly disproved by showing that distinct groups can share the same von Neumann algebra structure.
• Sphere packing received a new upper bound in high dimensions, which matters in geometry and information theory.
• Closest vector problem hardness results are relevant to lattice-based cryptography, including post-quantum security research.
• Quantum parallel repetition affects how repeated quantum games scale, which is central to theoretical quantum information.

How the announcement is being framed

The announcement is being described as a major research milestone because it combines original mathematical claims with formal proof verification. The release includes a long manuscript, model-generated reasoning summaries, and Lean certificates, which is more than a typical AI benchmark-style claim.

At the same time, the results are still best treated as claims until they are broadly examined by experts in the relevant fields. For high-impact mathematics, independent scrutiny is the standard test of whether a proof is accepted by the research community.

What this means for mathematics and AI

If the proofs hold up under expert review, Astra would represent a meaningful step toward AI systems that can contribute original, verifiable research rather than only summarize existing work. It would also show that formal proof assistants like Lean may become a central tool for checking advanced AI-generated mathematics.

For researchers, the deeper implication is not just that a model can propose answers, but that it can produce structured arguments in a form designed for verification. That combination of discovery and formal checking is what makes this announcement especially significant.

Conclusion

OpenAI’s Astra announcement is notable because it combines ambitious mathematical claims, cross-disciplinary scope, and machine-checked proofs in Lean. If the results withstand independent review, they could mark an important turning point in how AI contributes to frontier research.