Benchmarks › FrontierMath
93.7top score · GPT-6 Astra
69models ranked
13open-weights models
Accuracyhigher is better
Oct 20, 2026last updated
How the test works
1
Task
285 private problems (295 Tiers 1–3 minus 10 public)
- What the model doesSolve expert-written mathematics problems in FrontierMath Tiers 1–3 v2
- Kinds of tasksAdvanced mathematical problem solving
2
Run
1,000,000 cumulative input/output tokens
- Tools allowedPython execution, submit_answer
3
Score
Accuracy
- Graded byBinary verification of submitted Python answer; missing submission scores zero
- Better isHigher
- Test setPrivate evaluation; ten public Tiers 1–3 examples excluded
Sources: epoch.ai
FrontierMath Tiers 1–3 v2 leaderboardtop 10 of 69
Full leaderboard with filters →| Model | FrontierMath Tiers 1–3 v2 | Per task | Price | Speed | Supports | Context |
|---|---|---|---|---|---|---|
93.7max | $3.26 | $10 / $50 | 69 | 1.05M | ||
93.7max | $0.72 | $2 / $10 | 52 | 1.05M | ||
91.2max | $5.98 | $4 / $20 | 96 | 1M | ||
90.2max | $7.63 | $10 / $50 | 69 | 1M | ||
89.8max | $1.04 | $2 / $10 | 128 | 1.05M | ||
89.1max | $1.99 | $4 / $20 | 114 | 1.05M | ||
88.8max | $5.46 | $2 / $10 | 137 | 1M | ||
87.7xhigh | — | $30 / $180 | — | 1.05M | ||
87max | $8.75 | $10 / $50 | 67 | 1M | ||
86high | $1.99 | $2 / $10 | 84 | 1M |
Versions and boards on Model Menu
| Board | Models |
|---|---|
FrontierMath Tiers 1–3 v2Shown above | 69 |
| 54 | |
| 40 | |
| 33 | |
| 13 |
FrontierMath questions
What is FrontierMath?›
Solve expert-written mathematics problems in FrontierMath Tiers 1–3 v2. It has 285 private problems (295 Tiers 1–3 minus 10 public). Created by Epoch AI.
Which model is best on FrontierMath?›
GPT-6 Astra leads FrontierMath Tiers 1–3 v2 with 93.7, ahead of GPT-6.1 Sol (93.7) and Claude Opus 5.5 (91.2).
What is the best open-weights model on FrontierMath?›
Kimi K3 with 72.2, then GLM 5.3 (68.8).
How is FrontierMath scored?›
Accuracy; binary verification of submitted python answer; missing submission scores zero.
Which versions of FrontierMath are there?›
FrontierMath Tiers 1–3 v2 (69 models), FrontierMath Tier 4 v2 (54 models), FrontierMath (40 models), FrontierMath Tier 4 (33 models), FrontierMath Erdos (13 models).