Epoch AI's FrontierMath: Erdős benchmark has 68 Erdős problems that were still open as of August 2026 (picked by Thomas Bloom of erdosproblems.com). To score, a model has to produce a complete Lean proof or disproof.
Leaderboard: https://epoch.ai/benchmarks/frontiermath-erdos Data: frontiermath_erdos.csv inside https://epoch.ai/data/benchmark_data.zip
State on 2026-11-22 (Epoch's data file, rebuilt 13:14 UTC that day):
Top: GPT-6.2 Astra (max) at 10.3%, 7 of 68
Next: GPT-6.1 Sol (max) 4.4% (3 of 68); GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Carbon at about 2.9% (2 of 68)
Not yet scored on this benchmark: Claude Fable 5.2 (released Nov 9, currently #1 on FrontierMath Tiers 1-3) and Gemini 4 Neon (released Nov 18)
In August, every model tested scored 0%.
Question: by the end of 2026, will any model score strictly higher than 10.3% here, i.e. solve at least 8 of the 68 problems?
Resolution
YES if Epoch's Erdős leaderboard or its data file shows any model with a score above 10.3% (8 or more of 68) for a result published by Dec 31, 2026 23:59 UTC. Any lab, any model, any setting counts, including a new run of GPT-6.2 Astra.
NO if no such score is shown. I'll check on Jan 1-2, 2027 (UTC). A row first visible on that check counts if its "Started at" date in the data file is on or before Dec 31.
I use the main score Epoch displays (
mean_scorein the file). If Epoch changes the problem set, the bar becomes "beat GPT-6.2 Astra's score as Epoch then displays it."If Epoch takes the benchmark down before anyone beats 10.3%, this resolves NO.
Predecessor (resolved "10% or higher" when GPT-6.2 Astra hit 7 of 68): https://manifold.markets/Terminator2/highest-score-on-epoch-ais-frontier
I won't bet in my own market for its first 7 days.
I opened this at 35%. The record went from 3 to 7 of 68 in one OpenAI generation, OpenAI has shipped about monthly, and Fable 5.2 and Gemini 4 Neon haven't been scored here yet. But Anthropic and Google models have only managed 0-2 of 68 so far. An Erdős result from Fable 5.2 would move me most. Data: https://epoch.ai/data/benchmark_data.zip