Epoch AI's FrontierMath Erdős benchmark has 68 Erdős problems that were still open as of August 2026, picked by Thomas Bloom (erdosproblems.com). A model has to write a complete Lean proof or disproof. Leaderboard: https://epoch.ai/benchmarks/frontiermath-erdos
State on 2026-10-28 (Epoch's downloadable data, frontiermath_erdos.csv in https://epoch.ai/data/benchmark_data.zip):
Top score: GPT-6.1 Sol (max) at 4.4% (3 of 68)
Next: GPT-6 Astra and Claude Opus 5.5 at 2.9% (2 of 68)
In August, every model tested scored 0%.
Question: what is the highest score any model has on this leaderboard at close?
Resolution: On or shortly after Dec 1, 2026, I'll take the maximum score shown for any model on the Epoch FrontierMath Erdős leaderboard / in frontiermath_erdos.csv. Only runs whose evaluation started on or before Nov 30, 2026 (UTC, the CSV's "Started at" column) count. The score is the percentage Epoch shows (the CSV's best/mean score × 100). If the page and the CSV disagree, the CSV wins. If Epoch renames the benchmark, I'll use its direct successor. If they change the problem set itself, I'll go by the score on the original 68 problems where that can be told apart; if it can't, I'll explain in the comments before resolving.
Bucket edges: 4/68 = 5.9%, 5/68 = 7.4%, 6/68 = 8.8%, 7/68 = 10.3%.
1,000Answers start at 25% each. My numbers: below 5% about 55%, 5–7% about 23%, 7–10% about 14%, 10% or higher about 8%.
Why: the record went from 0 in August to 2 of 68 (GPT-6 Astra, early Sept) to 3 of 68 (GPT-6.1 Sol, late Sept), and nothing new has been added since. A single extra solved problem gets you to 5.9%. Gemini 4 Argon and GPT-6.1 Terra have no Erdős runs listed yet, and another big release in November would probably get tested within a few days. Each of these is a genuine open problem, though, so I expect any gains to come in ones and twos.
What would change my mind: a new model being listed at 4 or more out of 68, or Epoch reporting a new harness or scaffold for the benchmark. Data: https://epoch.ai/benchmarks/frontiermath-erdos