Benchmarks · Math
ArXivLean 09/2026 (MathArena)
Statements from September 2026 arXiv papers, formalized in the Lean proof language. The model has to prove them, and Lean checks the proof. Share solved. From MathArena.
10 models · 10 results · data as of 2026-11-18 · matharena.ai ↗
Leaderboard
GPT-6.2 Astra max
Heavy line: best score at each price. Thin lines join one model's effort levels. Cost uses each model's most subsidized plan, or API price with no plan data. 4 more models under All models in settings.
All results
10 rows| # | Model | Effort | Score | Subscription $ / task | API $ / task |
|---|---|---|---|---|---|
| 1 | GPT-6.2 Astra | max | 79.3% ±10.9 | $0.34 | $5.10 |
| 2 | GPT-6.2 Sol | max | 71.7% ±12.1 | $0.068 | $2.02 |
| 3 | Gemini 4 Carbon | high | 62.3% ±13.1 | $2.74 | $45 |
| 4 | Opus 5.5 | high | 47.2% ±13.4 | $0.18 | $14 |
| 5 | Gemini 4 Argon | high | 39.6% ±13.2 | $27 | $40 |
| 6 | GPT-6 Astra | max | 37.7% ±13.0 | $0.18 | $5.03 |
| 7 | GPT-6.1 Sol | max | 35.9% ±12.9 | $0.039 | $1.16 |
| 8 | Kimi K3.1 | max | 17.0% ±10.1 | $6.72 | $10 |
| 9 | DeepSeek V4.1 Flash | max | 5.7% ±5.9 | $0.13 | $1.87 |
| 10 | Qwen 3.8 | max | 1.9% ±2.8 | $9.00 | $9.00 |