OpenAI releases proofs for 372 major math problems, including Unique Games Conjecture

On October 6, 2026, OpenAI released a massive batch of 372 papers claiming to solve some of the most difficult open problems in mathematics and theoretical computer science. The release, guided by an advisory group including Timothy Gowers and Edward Witten, marks a potential turning point in how scientific discovery is conducted. Among the solved problems is the Unique Games Conjecture, a central question in complexity theory that has resisted proof for decades.

What happened

The announcement came as a shock to the academic community, particularly because the solutions were generated by an internal OpenAI model rather than a bespoke, multi-agent system. The model attempted approximately 8,000 problems and succeeded in solving about 5 percent of them, with each successful attempt requiring roughly three hours of GPT-Pro level compute. The results include not only the Unique Games Conjecture but also breakthroughs in derandomization, quantum complexity, and algorithmic efficiency.

The reaction among mathematicians has been mixed, ranging from excitement to disorientation. Dana Moshkovitz, a complexity theorist who has worked on the Unique Games Conjecture for her entire career, described the AI-generated proof as difficult to parse. She noted that the text feels disjointed, with irrelevant citations and bizarre constructions that do not align with traditional human intuition. Despite the confusion, the presence of Lean certificates for many of the proofs provides a layer of formal verification, even if human understanding lags behind.

This event highlights two emerging models for disseminating AI-driven scientific breakthroughs. OpenAI chose to release all papers publicly at once, creating a race for humans to digest and verify the results. In contrast, Anthropic recently collaborated with researchers Virginia Williams and Josh Alman to refine and announce their AI-assisted solutions to the 3SUM and All-Pairs Shortest Paths problems privately before public release. The OpenAI approach prioritizes transparency and speed, while the Anthropic model emphasizes curated communication and compensation for human collaborators.

How it works

The underlying technology is not a specialized supercomputer cluster but the latest internal iteration of OpenAI’s language model. It operates by generating proofs through probabilistic reasoning, leveraging vast amounts of training data on mathematical literature. The model does not merely retrieve existing answers but constructs novel arguments, such as a new type of error-correcting code for the Unique Games Conjecture that differs from known long or short codes.

Verification relies heavily on formal methods. Many of the 372 results include Lean certificates, which are machine-checkable proofs that ensure logical consistency. This allows the mathematical community to trust the correctness of the results even before fully comprehending the narrative flow of the arguments. The process shifts the role of human mathematicians from primary discoverers to interpreters and validators of AI-generated insights.

Key details

• OpenAI released 372 papers solving major open problems, including the Unique Games Conjecture and L=BPL.
• The solutions were generated by an internal model using about three hours of GPT-Pro compute per problem.
• Approximately 5 percent of the 8,000 attempted problems were successfully solved.
• Many proofs include Lean certificates for formal verification, though human-readable explanations remain scarce.
• The Unique Games Conjecture proof involves a new, recursive code construction that experts describe as alien and unintuitive.
• Cryptography problems were notably absent from the release, though investigations into cryptographic vulnerabilities are reportedly underway.

Why it matters

For software engineers and researchers, this development signals a shift in the nature of technical work. The ability of AI to solve problems that have stumped humans for decades suggests that future breakthroughs may come from interpreting machine output rather than deriving results from first principles. This changes the skill set required for high-level research, emphasizing verification, integration, and explanation over raw derivation.

The release also raises questions about the sustainability of traditional academic credit systems. If AI can generate proofs faster than humans can read them, the value of individual authorship diminishes. The community must now decide how to attribute credit and maintain motivation when the "heroic adventure" of discovery is partially automated. As Scott Aaronson noted, the challenge is to nurture a community that still cares about finding paths up these mountains, even when a teleportation machine exists.

What you can do

• Review the Lean certificates for any proofs relevant to your field to verify correctness independently.
• Focus on developing skills in formal verification tools like Lean to bridge the gap between AI output and human understanding.
• Collaborate with peers to digest and explain complex AI-generated proofs, sharing the workload of interpretation.
• Stay updated on the ethical and attribution guidelines emerging from institutions like the Simons Institute regarding AI-assisted research.
• Explore the implications of solved problems like L=BPL for your own work in algorithms and complexity.
• Engage in discussions about the future of mathematical credit and publication models in an AI-dominated landscape.