CRUXEval: leaderboard
Code reasoning benchmark with 800 Python functions. Tests input prediction (given output, predict input) and output prediction (given input, predict output). Simple functions but tricky reasoning.
Metric: Output Prediction pass@1 (%). Source: crux-eval.github.io. Saturation forecast: Estimated already saturated. 43 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 70 | #343 |
| 2 | GPT-4 (0613) | 68.7 | #491 |
| 3 | GPT-4 Turbo | 67.7 | #404 |
| 4 | Claude 3 Opus | 65.8 | #571 |
| 5 | GPT-3.5 Turbo (0613) | 49.4 | #1118 |
| 6 | starcoder2-15B | 47.1 | #1495 |
| 7 | Mixtral 8x7B | 40.5 | #1143 |
| 8 | starcoder2-7B | 36 | #1572 |
| 9 | Mistral 7B | 34.3 | #1397 |
| 10 | starcoder2-3B | 34.2 | #1596 |
| 11 | Phi-2 | 33.5 | #1511 |
| 12 | Phi-1.5 | 27.5 | #1549 |
| 13 | Stable-Code 3B | 26.7 | #1595 |
| 14 | Phi-1 | 21.7 | #1639 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cruxeval · How It Works · Data refreshed daily, snapshot 2026-10-20.