Background
<cite index="4-2">OpenAI Group PBC revealed on Saturday, August 2, that an internal version of Astra, the model family it calls its next major release, produced new results for ten problems in mathematics and theoretical computer science that had been open for at least a decade.</cite> <cite index="22-1">The announcement was embedded in a blog post titled "Ten advances in mathematics and theoretical computer science," which noted the results "were achieved by an internal version of Astra, our next major model."</cite> Astra has not been released publicly and carries no confirmed launch date or pricing.
The Proof Artifacts
<cite index="5-3">Alongside the announcement, OpenAI released a 249-page manuscript and Lean 4 proof certificates on GitHub under an Apache 2.0 license; the repository's "sorry" count stands at zero, indicating that every step across all ten formalized proofs is fully verified.</cite> <cite index="14-2">Human researchers turned the model's output into publishable papers, though OpenAI said the mathematical arguments themselves came from Astra.</cite>
Scope of the Results
<cite index="25-6">According to OpenAI, the Astra system generated mathematical arguments for ten long-standing problems spanning high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, cryptography, and extremal combinatorics.</cite> <cite index="5-4">Chief among the findings is an explicit construction of a non-sofic group, settling a question that has gone unanswered since Mikhail Gromov laid out the concept of soficity in 1999.</cite> <cite index="3-2">In cryptography, Astra generated a hardness-of-approximation result for the closest vector problem, relevant to post-quantum systems.</cite> <cite index="3-3">In combinatorics, it solved two classic Erdős problems: problem 183 on superexponential lower bounds for multicolor Ramsey numbers, and problems 146 and 180 on compactness and degeneration conjectures in extremal graph theory.</cite>
Compute Cost and Process
<cite index="14-4">OpenAI put the token cost for all ten solutions at roughly $2,000 at GPT-5.6 Sol application programming interface (API) rates.</cite> <cite index="6-5">The $2,000 figure covers the successful runs rather than every attempt the model made, making it a cost of publication rather than a total cost of discovery.</cite> <cite index="14-1">OpenAI describes Astra as a model family built to run long tasks by coordinating multiple agents over extended periods, an extension of the test-time reasoning work associated with research scientist Noam Brown, who called the results "a major step for scientific reasoning."</cite>
Context and Precedent
<cite index="2-4">AI for mathematics has progressed from solving eighth-grade-level math in 2023, to acing high-school-level math benchmarks in 2024, to achieving gold-level performance at the International Mathematical Olympiad (IMO) in 2025, and now to solving novel mathematical open questions in 2026.</cite> <cite index="3-10">In May 2026, OpenAI also reported using an unreleased model to refute the Erdős unit distance conjecture, and that result, according to the company, inspired subsequent developments by independent researchers in discrete geometry and communication complexity.</cite>
<cite index="1-8">The August 2026 Astra announcement differs from a disputed October 2025 GPT-5 claim in three structural ways: the problems span six distinct mathematical domains rather than a single catalogue; the proofs are formalized in Lean 4 with machine-checkable certificates that any reader can verify; and mathematician Thomas Bloom — the researcher who publicly refuted the 2025 claim — called this result "big news" and rated it more significant than OpenAI's May 2026 Erdős unit distance result.</cite>
Caveats and Open Questions
<cite index="8-5">The results are Lean-verified and have been reviewed informally by mathematicians who saw preprints, but none have yet gone through a formal, refereed journal process.</cite> <cite index="6-6,6-7">Outside researchers have noted that OpenAI staff helped prepare the papers and formalize the arguments, and because nobody outside the company can run the model that did the work, the result cannot be reproduced independently — only checked.</cite> <cite index="23-6">OpenAI's decision to publish machine-checkable Lean proofs directly addresses a key objection from the mathematical community regarding verifiability of AI-generated results, though questions raised by the Leiden Declaration — endorsed by the International Mathematical Union — regarding peer review, attribution, and consent remain unresolved.</cite> <cite index="20-15">The ten-proof dossier illustrates what OpenAI wants Astra to represent: a move from models that explain existing knowledge toward systems that can propose new knowledge and help formalize it.</cite>