AIME / MathArena
How the test works
Task
12 Apex 2025 problems
- What the model doesSolve twelve hard final-answer problems selected from 2025 mathematics competitions
- Kinds of tasksMathematical problem solving
- In → outText → Answer
Run
16 runs per task
- Agent / harnessDirect prompting; separately labelled self-verification agent uses four runs
- How runs combineMean success across problems and attempts; (Agent) GPT-5 uses four top-level runs
Score
Mean accuracy across attempts
- Graded byExtract boxed final answer and check mathematical correctness
- Better isHigher
MathArena · Apex leaderboardtop 10 of 34
Full leaderboard with filters →| Model | MathArena · Apex | Per task | Price | Speed | Supports | Context |
|---|---|---|---|---|---|---|
81max | $4.08 | $5 / $25 | 66 | 1M | ||
80xhigh | $2.63 | $5 / $30 | 113 | 1.05M | ||
70xhigh | — | $30 / $180 | — | 1.05M | ||
66unstated | — | $3 / $15 | — | 1M | ||
61unstated | — | $2 / $12 | — | 1M | ||
54xhigh | — | $2.50 / $15 | 127 | 1.05M | ||
41xhigh | — | $5 / $25 | — | 1M | ||
34high | — | $5 / $25 | 45 | 1M | ||
28max | — | $1.32 / $3.96 | — | 1M | ||
27max | — | $0.30 / $1.20 | — | 1M |
Versions and boards on Model Menu
| Board | Models |
|---|---|
MathArena · ApexShown above | 34 |
| 30 | |
| 24 | |
| 24 | |
| 24 | |
| 24 | |
| 22 | |
| 20 |
AIME / MathArena questions
What is AIME / MathArena?›
Solve twelve hard final-answer problems selected from 2025 mathematics competitions. It has 12 Apex 2025 problems. Created by SRI Lab at ETH Zurich and INSAIT.
Which model is best on AIME / MathArena?›
Claude Opus 4.8 leads MathArena · Apex with 81.3, ahead of GPT-5.5 (80.2) and GPT-5.4 Pro (69.8).
What is the best open-weights model on AIME / MathArena?›
Kimi K3 with 65.6, then DeepSeek V4 Pro 0813 (28.1).
How is AIME / MathArena scored?›
Mean accuracy across attempts; extract boxed final answer and check mathematical correctness; 16 runs per task per task.
Which versions of AIME / MathArena are there?›
MathArena · Apex (34 models), MathArena · Apex Shortlist (30 models), MathArena · ArXivMath June 2026 (24 models), MathArena · BrokenArXiv June 2026 (24 models), MathArena · BRUMO 2025 (24 models), MathArena · SMT 2025 (24 models), MathArena · CMIMC 2025 (22 models), MathArena · ArXivMath February 2026 (20 models).