Epoch AI - Frontiermath Erdos: leaderboard
Sixty-eight conjectures covering sixty-five problems posed or studied by Paul Erdős, all still open as of August 2026, curated by Thomas Bloom and formalised in Lean 4 against Mathlib. A model gets one attempt per conjecture with a $300 budget and 72 hours, a Lean toolchain, SageMath and an offline archive of 476,000 mathematics papers, and its proof has to survive Epoch’s Comparator checker in an isolated container. Bloom puts the prior art at three to five Erdős problems of this calibre ever solved by AI, which is why the board still reads in single digits.
Metric: Score. Source: epoch.ai. Saturation forecast: Around September 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 2.94 | #6 (GPT-6 Astra) |
| 2 | Claude Opus 5.5 (Max) | 2.94 | #3 (Claude Opus 5.5) |
| 3 | Claude Sonnet 5.5 (Max) | 2.9 | #11 (Claude Sonnet 5.5) |
| 4 | GPT-5.5 (xHigh) | 0 | #31 (GPT-5.5) |
| 5 | GPT-5.6 Sol (Max) | 0 | #19 (GPT-5.6 Sol) |
| 6 | Claude Fable 5.1 (Max) | 0 | #9 (Claude Fable 5.1) |
| 7 | Claude Fable 5 (Max) | 0 | #16 (Claude Fable 5) |
No result here: #7 GPT-6.1 Sol, #8 Gemini 4 Argon, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=epoch-ai-frontiermath-erdos · How It Works · Data refreshed daily, snapshot 2026-10-20.