In a decisive pivot toward academic transparency, OpenAI publicly released `github.com/openai/math` on October 6, 2026, unveiling 722 mathematical manuscripts generated by an advanced internal reasoning model across 372 distinct 'result families.'
The release marks a deliberate departure from the controversial era of unverified promotional announcements that previously drew sharp rebuke from academic mathematicians. To design the release framework, OpenAI convened and consulted with an independent Advisory Group on Mathematics and Artificial Intelligence hosted at the Institute for Advanced Study (IAS) in Princeton. The advisory committee advised that AI laboratories must accompany mathematical assertions with rigorous machine-verifiable artifacts and transparent compute logs.
The repository satisfies these criteria by adopting a dual-track exposition methodology. For each mathematical topic, OpenAI provides both a standard human-readable LaTeX manuscript detailing the intuitive strategy and algebraic lemmas, and a formal proof script compiled in the Lean 4 interactive theorem prover. This ensures that mathematicians can verify the logical soundness of candidate theorems independently without relying on the credibility or reputational authority of the model.
The breadth of mathematical problems addressed in the corpus is expansive. Among the 372 result families are candidate lemmas tightening approximation ratios for the Unique Games Conjecture in computational complexity, new non-trivial zero-free regions for Dirichlet L-functions relating to the Generalized Riemann Hypothesis, and structural decompositions of the Hodge conjecture for CM abelian varieties.
Critically, OpenAI included comprehensive inference telemetry. Generating each result family required an average of three hours of continuous 'thinking' compute—equivalent to hundreds of thousands of reasoning tokens evaluating branching proof trees, backtracking when encountering dead ends, and compiling intermediate formal statements in an embedded Lean kernel.
Initial responses from the international mathematics community have praised the methodology. By replacing marketing press releases with open-source Git repositories containing verifiable Lean syntax and reproduction scripts, OpenAI has established a rigorous benchmark for how synthetic artificial intelligence can contribute meaningfully to human foundational science.