On October 7, 2026, OpenAI’s mathematics repository history recorded three manuscript withdrawals after a sign error, proof repairs to 14 other manuscripts, and citation or version updates to 13 more. The current catalog lists 719 manuscripts in 372 result families, down from 722 at the October 6 launch.
OpenAI announced the collection on October 6, describing it as work produced largely by an unreleased internal model. NeoTeo previously covered a separate OpenAI mathematics claim involving Navier–Stokes in an earlier report.
OpenAI records withdrawals and manuscript revisions
OpenAI’s October 7 history says a sign error invalidated an argument behind three manuscripts, which the company withdrew. The same history lists proof repairs in 14 other manuscripts and citation or version updates in 13 additional ones. These are separate revision categories, not further withdrawals.
The catalog went from 722 manuscripts to 719
The launch catalog contained 722 manuscripts; after the three withdrawals, OpenAI’s repository listed 719. The count of result families remained 372 in both snapshots.
A family groups related documents. It can include a main result alongside companion arguments, consequences, or alternative proofs, so 372 families do not mean 372 individual papers—or 372 independently accepted theorems. OpenAI also says its internal model was posed approximately 4,000 problems. That evaluation figure is not a problem-by-problem tally of the published catalog: OpenAI describes the release as a curated collection.
What OpenAI’s Lean tally measures
OpenAI reported that 300 of the 719 top-line results were formalized in Lean, about 42%. Lean is a proof assistant: software that checks a proof written in a formal system against its encoded statement and assumptions.
That tally measures formalization, not independent acceptance by mathematicians. OpenAI says the collection’s results are at different verification stages and that some results without Lean formalizations could contain issues.
Mathematical review remains distinct from formalization
A Lean check applies to a formal proof and the statement encoded for it. Whether that statement captures the intended result, and how the result stands up to mathematical scrutiny, are separate questions. OpenAI’s repository reports differing verification stages across the collection; its 300 formalized results are not a count of papers accepted through independent review.
