EvoEval Subtle: leaderboard
The Subtle benchmark of EvoEval (arXiv) — problems with small but meaning-changing edits; Pass@1 (%).
Metric: Pass@1 (%). Source: evo-eval.github.io. Saturation forecast: Estimated already saturated. 51 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 Turbo | 82 | #404 |
| 2 | GPT-4 | 76 | #438 |
| 3 | ChatGPT | 70 | #900 |
| 4 | Magicoder-S-DS-6.7B | 67 | #1256 |
| 5 | Claude 3 Haiku | 65 | #803 |
| 6 | Claude 2 | 64 | #761 |
| 7 | Phind-CodeLlama-34B-v2 | 63 | #1589 |
| 8 | Qwen 1.5 72B | 61 | #753 |
| 9 | deepseek-coder-6.7B-instruct | 61 | #1335 |
| 10 | Magicoder-S-CL-7B | 58 | #1489 |
| 11 | Gemini 1.0 Pro | 53 | #933 |
| 12 | deepseek-coder-1.3B-instruct | 53 | #1585 |
| 13 | Phi-2 | 49 | #1511 |
| 14 | Qwen 1.5 14B | 48 | #1096 |
| 15 | Qwen-14B | 45 | #1171 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=evoeval-subtle · How It Works · Data refreshed daily, snapshot 2026-10-20.