News · 2026-09-22

OpenAI says an internal model resolved 100-plus math problems and asks an independent group to advise on disclosure

OpenAI says a new internal model has resolved more than 100 long-standing mathematics problems, and it has established an independent advisory relationship to help guide disclosure. The verified news is the claim and the governance structure, not a publicly checkable catalogue of 100 solutions.

Key facts

• OpenAI says it began training the relevant internal model on 28 August.
• It says that model “has now resolved more than 100 long-standing open problems” across most mathematical fields.
• The announcement supplies no list of those problems, manuscripts, Lean files, or release schedule.
• The independent Advisory Group on Mathematics and AI says it has no decision-making power at any AI company.

The phrase “resolved more than 100” needs careful handling. It is a real, dated OpenAI statement. It is not a public database of claims that readers, referees, or other mathematicians can inspect. The announcement does not identify the problems, say whether they are full proofs or partial advances, explain whether any overlap an earlier public package, name who has checked them, or provide a field-by-field breakdown. Reporting “OpenAI says” is warranted. Reporting that mathematics has received 100 accepted solutions is not.

OpenAI's practical response is to work with the Advisory Group on Mathematics and AI. Its initial members include François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, and Melanie Matchett Wood. In a guest post, Terence Tao explains that the group formed independently after OpenAI approached some mathematicians about an external board. It will publish recommendations and take input from the mathematical community. OpenAI says the group will advise on significance, review and communication, dissemination, professional standards, and tools for research.

What it cannot do is equally important. The group is not a journal referee panel, a prize committee, or a body that can veto an AI company's release. Its statement says it “has no decision-making power at any AI company,” and OpenAI says it will not advise on the pace of internal mathematics progress. Think of it as a standards and translation layer between a laboratory that may generate a great many technical claims and a field that has to understand, attribute, teach, and check them. It can improve the handoff; it cannot make unpublished evidence reviewed by fiat.

There are already more concrete, separate public records. OpenAI's August package of ten advances names ten results and says the system formalized arguments in Lean. Its Navier–Stokes page publishes a claimed result and a Lean formalization produced through a large coordinated agent system. Those artifacts are not the same as the September 100-plus statement. The new announcement does not say whether the count includes the earlier ten or represents only later work.

The Navier–Stokes example shows why labels matter. OpenAI says it proved finite-time singularity under a stated setting and does not claim the Clay prize. Clay Mathematics Institute said the problem had “apparently been settled,” but its rules require publication in a qualifying outlet, at least two years, and general acceptance before a prize is considered. A model output, a Lean-checked formalization, a peer-readable proof, and a prize resolution are different stages of the same long road.

The advisory group arises amid the mathematicians' declaration that warned against treating famous open problems as simple capability benchmarks. Its core concern is not that a computer must not find results. It is that mathematics becomes durable only through explanation, independent understanding, attribution, and teaching. A flood of answers without that social machinery can make the field harder, not easier, to navigate.

The strongest counterargument is that critics may be asking an unusually capable research system to wait for institutions that were built for slower discovery. That tension is real. Yet the absence of a list means the public cannot judge scope, novelty, correctness, or overlap. The important development is a new governance pattern: OpenAI is signaling discovery capacity ahead of normal disclosure, while leading mathematicians are trying to construct a process that distinguishes a company claim from a community result.

A responsible next disclosure would connect each result to a precise statement, a human-readable proof, a formal artifact where possible, prior literature, named scrutiny, and an explicit status. Until then, the meaningful number is zero publicly itemized claims in this new batch.

That is not a criticism of formal methods; it is a description of the evidence readers need before a broad research claim becomes durable public knowledge.

Primary source, verified: read the paper →

Ground Truth. (2026, September 22). OpenAI says an internal model resolved 100-plus math problems and asks an independent group to advise on disclosure. Ground Truth. https://groundtruth.day/news/openai-math-advisory-group-hundred-claims.html

@misc{groundtruth:openai-math-advisory-group-hundred-claims,
  title  = {OpenAI says an internal model resolved 100-plus math problems and asks an independent group to advise on disclosure},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/openai-math-advisory-group-hundred-claims.html}
}

Topics: mathematics · scientific-ai · governance · formal-verification · research