FrontierMath - Tier 4: leaderboard
Research-level mathematics problems at the frontier of active mathematical inquiry, requiring expertise beyond postdoctoral level. The 48 hardest unpublished problems in Epoch AI's FrontierMath benchmark.
Metric: Accuracy (%, 48 problems). Source: epoch.ai. Saturation forecast: Around March 2027. 72 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 Pro (xHigh) | 39.6 | #17 (GPT-5.5 Pro) |
| 2 | GPT-5.4 Pro (xHigh) | 37.5 | #32 (GPT-5.4 Pro) |
| 3 | GPT-5.5 (xHigh) | 35.4 | #31 (GPT-5.5) |
| 4 | GPT-5.2 Pro | 31.3 | #61 |
| 5 | Claude Opus 4.8 (Max) | 31.25 | #37 (Claude Opus 4.8) |
| 6 | GPT-5.4 (xHigh) | 27.1 | #79 (GPT-5.4) |
| 7 | Claude Opus 4.7 (xHigh) | 22.92 | #49 (Claude Opus 4.7) |
| 8 | Claude Opus 4.6 (Max) | 22.9 | #63 (Claude Opus 4.6) |
| 9 | GPT-5.2 (High) | 18.8 | #110 (GPT-5.2) |
| 10 | GPT-5.2 (xHigh) | 18.8 | #110 (GPT-5.2) |
| 11 | Gemini 3 Pro (Preview) | 18.75 | #66 |
| 12 | Gemini 3.1 Pro (Preview) | 16.7 | #58 |
| 13 | GPT-5.2 (Medium) | 16.7 | #110 (GPT-5.2) |
| 14 | Muse Spark | 14.6 | #64 |
| 15 | Kimi K2.6 | 14.58 | #104 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=frontiermath-tier-4 · How It Works · Data refreshed daily, snapshot 2026-10-20.