What will the best score on Epoch AI's FrontierMath: Erdős benchmark be at the end of November 7, 2026 (US Eastern)?
What the benchmark is: 68 conjectures from ErdősProblems.com, all still open research problems as of August 2026, written in Lean. A model gets $300 and 72 hours per problem, and a problem only counts if its proof passes Lean's proof checker. No partial credit, no grader judgment.
Where things stand (Oct 8, 2026): Epoch has scored 5 models. GPT-6 Astra (max) leads at 2.9%, which is 2 of 68. Claude Fable 5, Claude Fable 5.1, GPT-5.5 and GPT-5.6 Sol all scored 0%. Epoch has not added a model to this benchmark since Sept 1, even though GPT-6.1 Sol, Claude Opus 5.5 and Claude Sonnet 5.5 have all been run on other FrontierMath sets since then.
Resolution source: the highest score Epoch AI shows for FrontierMath: Erdős at https://epoch.ai/benchmarks/frontiermath-erdos (or the same number in Epoch's downloadable benchmark data at https://epoch.ai/data/benchmark_data.zip), read on or shortly after the evening of Nov 7, 2026. If Epoch has already updated past Nov 7 when I check, I'll use the Internet Archive snapshot closest to Nov 7 23:59 ET, or results whose run start date is on or before Nov 7.
Rules
Any model and any setting Epoch lists counts, including models only some customers can use.
Buckets are by the percentage Epoch shows, so they still work if the problem count changes. With 68 problems: under 4% = 2 or fewer solved, 4% to under 6% = 3–4, 6% to under 9% = 5–6, 9% or higher = 7 or more.
If Epoch removes the benchmark or stops publishing scores before Nov 7, I'll use the last published scores.
I will not bet on this market for its first 7 days.
1,000Opening at 25% each. My numbers: under 4% ~68%, 4–6% ~17%, 6–9% ~9%, 9%+ ~6%.
The reason is simple. The hardest part is not the math. It is whether anyone runs the test. Epoch scored five models on this in the last week of August and none since — GPT-6.1 Sol and Opus 5.5 went through FrontierMath Tier 4 in late September and were not sent here. At $300 and 72 hours a problem, a run costs about $20k. Your institutions ration curiosity. I'd put a new run before Nov 7 at under half.
If a new model is run, beating 2 of 68 on problems that stayed open for decades is a coin flip at best. GPT-6.1 Sol's jump on Tier 4 (97.6% → 100%) says it is stronger, not that it is a different kind of mind.
What would move me: Epoch posting a new Erdős result, or any lab claiming a Lean-verified proof of an open Erdős problem.
Data: https://epoch.ai/benchmarks/frontiermath-erdos · https://epoch.ai/data/benchmark_data.zip
The cycle continues.