KernelBench Hard: leaderboard
Metric: Mean Peak Fraction Across Problems (ratio). Source: benchmarklist.com. 59 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Fable 5 (Max) | 0.35 |
| 2 | Qwen 3.8 Max (xHigh) | 0.34 |
| 3 | GLM-5.2 | 0.26 |
| 4 | GLM-5.3 | 0.25 |
| 5 | Claude Opus 4.8 | 0.24 |
| 6 | DeepSeek V4 Pro | 0.24 |
| 7 | Claude Opus 4.8 (Max) | 0.24 |
| 8 | Grok 4.6 (xHigh) | 0.24 |
| 9 | Claude Opus 5 (Max) | 0.23 |
| 10 | Kimi K3 | 0.23 |
| 11 | MiniMax-M3 | 0.22 |
| 12 | GPT-5.6 Sol (xHigh) | 0.22 |
| 13 | Claude Fable 5 | 0.22 |
| 14 | DeepSeek V4 Flash (0731) | 0.2 |
| 15 | LongCat 2.0 | 0.16 |
Interactive version: theaggregate.ai/benchmark?slug=kernelbench-hard · How It Works · Data refreshed daily, snapshot 2026-09-19.