Epoch AI introduces FrontierMath Erdős to measure AI math capabilities

Epoch AI introduces FrontierMath Erdős to measure AI math capabilities

Epoch AI published this announcement on September 1, 2026, introducing FrontierMath Erdős as a more rigorous way to use Erdős problems to track AI math capabilities. Mathematician Thomas Bloom selected problems he considered significant and difficult; Epoch AI acknowledges that this curation is subjective. The benchmark uses Lean, a specialized programming language for expressing mathematical proofs, to check solutions programmatically, and makes its code and problem list public to support transparency and replicability. In the fixed-protocol run, only GPT-6 Astra solved problems, producing verified results on 2 of the 68; Epoch AI says this reflects a particular scaffold and budget, not the chance of solving any arbitrary open problem. Separate, less systematic attempts produced additional solutions but do not count toward the benchmark score. Across all attempts, they cost over $220,000 in compute, compared with roughly $20,000 for the benchmark run. Epoch AI describes breakthroughs on this list as real but still fairly uncommon. It says Erdős problems are not a sample of all mathematics, and concludes that mathematicians must examine the nature of the solutions before their significance can be assessed.

Publication: 63543660-4cd6-d9ff-7adf-e4a267e1634d

attribution — pass: This is a single Epoch AI publisher-page announcement, not a listing. The reading attributes the benchmark design, Bloom's curation, results and conclusions to Epoch AI or Bloom, and does not present them as established fact.

With our new math benchmark, FrontierMath Erdős, we aim to add some rigor when it comes to using Erdős problems to measure AI math capabilities.

summary — uncertain: The reading is mostly clear, and Lean is explained. However, it reframes the conclusion as "reflects a particular scaffold and budget, not the chance of solving any arbitrary open problem." That reads as a disclaimer and obscures the source's actual point: AI does not yet have a high chance of solving an arbitrary problem of this caliber, at least when the result must be formalized within $300. "Scaffold" is also unexplained. Smallest repair: state that, in this scaffold at this budget, the result suggests AI's chance on an arbitrary problem of this caliber is not yet high.

it is not yet the case that AI has a particularly high chance of solving an arbitrary such problem — at least not if it has to formalize the result within a budget of $300 per problem

date — pass: The reading gives September 1, 2026, which is the UTC calendar day of the instant originalPublished. The source's August 2026 references are consistent with that date. Forecasts, such as waiting for mathematicians' commentary, are not presented as outcomes, and no hindsight is used.

no solution to any of the 68 problems was known as of August 2026

support — uncertain: Most claims match the source: 68 problems, 2 of 68 solved in the benchmark run, 5 solved across all attempts, over $220,000 versus roughly $20,000, and "real but still fairly uncommon." Two passages drift. First, "any arbitrary open problem" broadens the source's "arbitrary such problem" and drops the $300 formalization condition. Second, saying Epoch AI "concludes that mathematicians must examine the nature of the solutions before their significance can be assessed" overstates the source, which says only that the solutions' nature will be informative and that it will wait for mathematicians' comments.

The nature of the AI solutions to these five problems will also be informative. For that, we’ll have to wait for mathematicians to digest and comment on the results.

media — pass: No media was selected and none was supplied or verified, so nothing is misidentified or irrelevant.