OpenAI posts 372 math results from one agent and one prompt, and keeps the prompt

OpenAI posted 372 mathematical results from an unreleased internal model to a GitHub repository this week, each presented as solving an open problem or making substantial progress on one.^{1 The Decoder 2026-10-07 372 results on GitHub; one prompt, one agent, about three hours Pro compute each; many with Lean proofs; IAS advisory group limited to communication; 25 Fields medallists signed an open letter objecting to mass production. Open source} Nearly every result came from a single prompt handed to a single agent, sometimes over several attempts, at an average of about three hours of ChatGPT Pro Thinking compute per result.^{1 The Decoder 2026-10-07 372 results on GitHub; one prompt, one agent, about three hours Pro compute each; many with Lean proofs; IAS advisory group limited to communication; 25 Fields medallists signed an open letter objecting to mass production. Open source} The model attempted roughly 4,000 problems, with an acceptance rate around 18 percent.^{3 AI News Weekly 2026-10-07 About 4,000 problems, about 18 percent accepted; GitHub Issues disabled and prompts withheld; IAS group had asked for model, prompt and per problem compute; Sutherland urges treating claims as unverified. Open source}

Among the results is a claimed proof of Subhash Khot's Unique Games Conjecture, one of the central open questions in the theory of approximation, accompanied by a Lean certificate.^{2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source} If it holds, it settles the hardness of approximating a family of optimisation problems that complexity theorists have organised a research programme around for two decades.^{2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source}

What the release is and is not

Three facts constrain how far to take the headline. First, verification is partial. Many results include Lean formalisations, but not all, and Scott Aaronson writes that no human has yet understood most of the proofs.^{1 The Decoder 2026-10-07 372 results on GitHub; one prompt, one agent, about three hours Pro compute each; many with Lean proofs; IAS advisory group limited to communication; 25 Fields medallists signed an open letter objecting to mass production. Open source 2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source} A Lean certificate shows the formal statement follows from the axioms; it does not show the formal statement says what the paper claims it says. Second, the batch has already had at least one public retraction: a Hodge conjecture result withdrawn over a sign error.^{2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source} Third, the method is not reproducible from outside. OpenAI withheld the prompts and disabled GitHub Issues on the repository, and released only average compute figures.^{3 AI News Weekly 2026-10-07 About 4,000 problems, about 18 percent accepted; GitHub Issues disabled and prompts withheld; IAS group had asked for model, prompt and per problem compute; Sutherland urges treating claims as unverified. Open source}

The single agent framing matters because it is a claim about cost. An earlier Navier Stokes result reportedly used a swarm of 10,000 agents and millions of dollars in compute.^{1 The Decoder 2026-10-07 372 results on GitHub; one prompt, one agent, about three hours Pro compute each; many with Lean proofs; IAS advisory group limited to communication; 25 Fields medallists signed an open letter objecting to mass production. Open source} Three hours of Pro compute per accepted result puts serious open problems within reach of an ordinary paid subscription, if and when the model ships.^{3 AI News Weekly 2026-10-07 About 4,000 problems, about 18 percent accepted; GitHub Issues disabled and prompts withheld; IAS group had asked for model, prompt and per problem compute; Sutherland urges treating claims as unverified. Open source}

The process fight

The more consequential argument is about who decides the pace. OpenAI consulted the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, but mathematicians could advise only on how the results were communicated, not on whether or how fast they were produced.^{1 The Decoder 2026-10-07 372 results on GitHub; one prompt, one agent, about three hours Pro compute each; many with Lean proofs; IAS advisory group limited to communication; 25 Fields medallists signed an open letter objecting to mass production. Open source} The group had recommended publishing the model, the exact prompt and the compute time per problem; OpenAI released averages instead.^{3 AI News Weekly 2026-10-07 About 4,000 problems, about 18 percent accepted; GitHub Issues disabled and prompts withheld; IAS group had asked for model, prompt and per problem compute; Sutherland urges treating claims as unverified. Open source}

Twenty five Fields medallists have signed an open letter objecting to mass production of results, arguing that problem solving is a proxy for understanding and that a flood of true statements could damage the field.^{1 The Decoder 2026-10-07 372 results on GitHub; one prompt, one agent, about three hours Pro compute each; many with Lean proofs; IAS advisory group limited to communication; 25 Fields medallists signed an open letter objecting to mass production. Open source} Andrew Sutherland of MIT urged treating the single agent claims as unverified until the model is released and results can be replicated.^{3 AI News Weekly 2026-10-07 About 4,000 problems, about 18 percent accepted; GitHub Issues disabled and prompts withheld; IAS group had asked for model, prompt and per problem compute; Sutherland urges treating claims as unverified. Open source}

We assess with high confidence that this is now the central tension in AI for mathematics. The capability is no longer the open question. The open question is whether the field's verification and refereeing capacity can absorb output at this rate, and whether a company that controls the model, the prompt and the release schedule can also be the arbiter of what counts as a result.

Second order effects

Refereeing becomes the bottleneck. Hundreds of claimed results, most of them unread by humans, land on a community whose review capacity is measured in dozens of careful readings a year.^{2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source}

Formal verification becomes a gate rather than a courtesy. Results without Lean certificates will be treated as lower grade, and pressure grows to formalise the rest.^{1 The Decoder 2026-10-07 372 results on GitHub; one prompt, one agent, about three hours Pro compute each; many with Lean proofs; IAS advisory group limited to communication; 25 Fields medallists signed an open letter objecting to mass production. Open source}

Disclosure norms harden. The IAS group's recommendation of model, prompt and per problem compute is a concrete standard, and its rejection makes it a public benchmark against which the next release will be judged.^{3 AI News Weekly 2026-10-07 About 4,000 problems, about 18 percent accepted; GitHub Issues disabled and prompts withheld; IAS group had asked for model, prompt and per problem compute; Sutherland urges treating claims as unverified. Open source}

Who gains and who loses

Gains: OpenAI, which has made the strongest public case yet for its unreleased model's reasoning, ahead of any commercial launch.^{2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source} Formal methods researchers, whose tools are now the accepted standard of trust.^{1 The Decoder 2026-10-07 372 results on GitHub; one prompt, one agent, about three hours Pro compute each; many with Lean proofs; IAS advisory group limited to communication; 25 Fields medallists signed an open letter objecting to mass production. Open source}

Loses: mathematicians whose research programmes target the problems the batch claims to settle, who now face either verification work or obsolescence on a timeline they did not choose. Journals, whose role as the arbiter of record is bypassed by a GitHub repository.^{1 The Decoder 2026-10-07 372 results on GitHub; one prompt, one agent, about three hours Pro compute each; many with Lean proofs; IAS advisory group limited to communication; 25 Fields medallists signed an open letter objecting to mass production. Open source}

The counter case

The case for OpenAI's approach is that results are results. A Lean certified proof does not need the prompt to be checked, and withholding a method is a normal commercial choice that does not make a verified theorem less true. On that view the IAS recommendations were a request for academic norms the company never agreed to, and a spokesperson has said as much.^{3 AI News Weekly 2026-10-07 About 4,000 problems, about 18 percent accepted; GitHub Issues disabled and prompts withheld; IAS group had asked for model, prompt and per problem compute; Sutherland urges treating claims as unverified. Open source} The weakness is that most of the batch is not yet verified by anyone outside the company, so the strongest version of the defence currently applies only to a subset.^{2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source}

Assessment

We assess with moderate confidence that a majority of the Lean certified results survive expert review, and with low confidence that the uncertified remainder does at the same rate, given one retraction already.^{2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source} We assess with moderate confidence that OpenAI does not publish the prompt before the model itself is released.^{3 AI News Weekly 2026-10-07 About 4,000 problems, about 18 percent accepted; GitHub Issues disabled and prompts withheld; IAS group had asked for model, prompt and per problem compute; Sutherland urges treating claims as unverified. Open source}

What to watch

First, independent confirmation or refutation of the Unique Games Conjecture proof from complexity theorists by the end of 2026.^{2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source}

Second, further retractions from the batch, and the count by year end.^{2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source}

Third, whether the model reaches paying ChatGPT customers within the next few months, as has been suggested, and whether the prompt ships with it.^{2 Scott Aaronson, Shtetl-Optimized 2026-10-07 Includes a Lean certified Unique Games Conjecture proof; about 4,000 problems attempted; most proofs not yet understood by humans; Hodge result retracted for a sign error; no cryptography in the batch. Open source}

Fourth, a formal response from journals or the Institute for Advanced Study group on standards for AI generated submissions before the end of the academic year.^{1 The Decoder 2026-10-07 372 results on GitHub; one prompt, one agent, about three hours Pro compute each; many with Lean proofs; IAS advisory group limited to communication; 25 Fields medallists signed an open letter objecting to mass production. Open source}