Benchmark · Math
FrontierMath Tiers 1-3 (v2)
accuracy % (%) Higher is better Benchmark homepage
What it measures
FrontierMath is a set of original, very hard mathematics problems written and vetted by expert mathematicians, maintained by Epoch AI. Version 2, released on 2026-06-12, corrected errors in 42% of problems; its base set, Tiers 1-3, has 295 problems ranging from advanced undergraduate to early-career research level. Epoch runs the evaluations itself on the private problem set: models may reason and run Python and must submit an answer() function, which is checked automatically.
FrontierMath was developed with funding from OpenAI, which has exclusive access to a subset of the problems. Scores from v1 and from Tier 4 (a separate 43-problem expansion set) are not comparable with Tiers 1-3 (v2). The best models score about 94%, so the benchmark separates mid-size models better than it separates frontier models. Epoch's numbers come from its own harness (Python tool, 1M-token limit), which can differ from providers' internal setups. Margins are about ±1.4 to ±2.6 points.
Results
| Model | Score | Setting | Reported by | Measured | Verified | Source |
|---|---|---|---|---|---|---|
| GPT-6 Astra OpenAI | 93.7 % | max effort; Epoch harness (Python tool) | Third party | — | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| GPT-6.1 Sol OpenAI | 93.7 % | max effort; Epoch harness (Python tool) | Third party | — | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Claude Opus 5.5 Anthropic | 91.2 % | Tiers 1-3 v2; max effort; Epoch harness (Python tool) | Third party | 2026-09-22 | 2026-10-12 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Claude Fable 5.1 Anthropic | 90.2 % | Tiers 1-3 v2; max effort; Epoch harness (Python tool) | Third party | 2026-09-01 | 2026-10-12 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| GPT-6 Sol OpenAI | 89.8 % | max effort; Epoch harness (Python tool) | Third party | — | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Claude Sonnet 5.5 Anthropic | 88.8 % | Tiers 1-3 v2; max effort; Epoch harness (Python tool) | Third party | 2026-09-29 | 2026-10-12 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Gemini 4 Argon Google | 86 % | Tiers 1-3 v2; high effort; Epoch harness (Python tool) | Third party | 2026-10-13 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Muse Wonder Meta | 82.1 % | Tiers 1-3 (v2); max effort; Epoch harness (Python tool) | Third party | 2026-10-12 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| GPT-6 Luna OpenAI | 78.9 % | max effort; Epoch harness (Python tool) | Third party | — | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Claude Haiku 5.5 Anthropic | 75.1 % | Tiers 1-3 v2; max effort; Epoch harness (Python tool) | Third party | 2026-10-07 | 2026-10-12 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Qwen3.8-Max Alibaba Cloud (Qwen) | 74.7 % | Tiers 1-3 v2; xhigh effort; Epoch harness (Python tool) | Third party | 2026-08-04 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Muse Spark 1.3 Meta | 74.4 % | Tiers 1-3 (v2); xhigh effort; Epoch harness (Python tool) | Third party | 2026-09-16 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Muse Spark 1.3 Meta | 74.04 % | Tiers 1-3 (v2); max effort; Epoch harness (Python tool) | Third party | 2026-09-18 | 2026-10-25 | Epoch AI benchmarking hub data download (benchmark_data.zip) |
| DeepSeek-V4.1-Flash DeepSeek | 71.2 % | snapshot DeepSeek-V4.1-Flash-1017; max effort; Epoch harness (Python tool) | Third party | 2026-10-19 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Muse Spark Open Meta | 69.8 % | Tiers 1-3 (v2); xhigh effort; Epoch harness (Python tool) | Third party | 2026-10-15 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Gemini 3.8 Flash Google | 68.4 % | Tiers 1-3 v2; high effort; Epoch harness (Python tool) | Third party | 2026-09-02 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| DeepSeek-V4.1-Flash DeepSeek | 67.4 % | max effort; Epoch harness (Python tool) | Third party | 2026-10-08 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Grok 4.6 xAI | 66 % | Tiers 1-3 v2; xhigh effort; Epoch-administered | Third party | — | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Qwen3.8-Max Alibaba Cloud (Qwen) | 65.6 % | Tiers 1-3 v2; snapshot qwen3.8-max-0902; xhigh effort; Epoch harness (Python tool) | Third party | 2026-09-02 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| DeepSeek-V4-Pro DeepSeek | 64.6 % | max effort; Epoch harness (Python tool) | Third party | 2026-08-19 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Gemini 3.1 Pro Google | 59.6 % | Tiers 1-3 v2; effort not shown; Epoch harness (Python tool) | Third party | 2026-06-11 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Gemini 3.6 Flash Google | 58.9 % | Tiers 1-3 v2; high effort; Epoch harness (Python tool) | Third party | 2026-08-02 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Grok 4.5 xAI | 57.2 % | Tiers 1-3 v2; high effort; Epoch-administered | Third party | — | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Grok 4.7 xAI | 53 % | Tiers 1-3 v2; xhigh effort; Epoch-administered | Third party | — | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Qwen3.7-Plus Alibaba Cloud (Qwen) | 34.4 % | Tiers 1-3 v2; no thinking; Epoch harness (Python tool) | Third party | 2026-08-29 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Gemini 3.1 Flash-Lite Google | 27.7 % | Tiers 1-3 v2; high effort; Epoch harness (Python tool) | Third party | 2026-08-30 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Gemini 3.5 Flash-Lite Google | 26 % | Tiers 1-3 v2; high effort; Epoch harness (Python tool) | Third party | 2026-08-02 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Claude Sonnet 4.5 Anthropic | 23.9 % | Tiers 1-3 v2; 32k thinking; Epoch harness (Python tool) | Third party | 2026-06-11 | 2026-10-12 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Gemini 3.1 Flash-Lite Google | 22.5 % | Tiers 1-3 v2; low effort; Epoch harness (Python tool) | Third party | 2026-08-30 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
| Gemini 3.1 Flash-Lite Google | 21.4 % | Tiers 1-3 v2; minimal effort; Epoch harness (Python tool) | Third party | 2026-08-28 | 2026-10-25 | FrontierMath Tiers 1-3 (v2) leaderboard (Epoch AI) |
Sorted by score, without rank numbers: each result was measured under its own setting (effort level, harness, tools, test version), as stated by the source. Scores are only directly comparable between rows with the same setting. Self-reported results come from the provider and have not been reproduced independently.