KernelBench Hard: leaderboard
KernelBench Hard is a text-only benchmark for autonomous coding agents on GPU kernel engineering tasks. It assesses correctness and speed against hardware baselines. The pinned metric, Mean Peak Fraction Across Problems, measures performance as a ratio relative to the problem peak across the suite. Source: BenchmarkList’s attributed result table.
Metric: Mean Peak Fraction Across Problems (ratio). Source: benchmarklist.com. Saturation forecast: Not forecast. 63 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.2 | 0.25 | #80 |
| 2 | DeepSeek V4 Pro | 0.23 | #112 |
| 3 | Claude Opus 4.8 | 0.22 | #39 |
| 4 | Grok 4.6 (xHigh) | 0.22 | #29 (Grok 4.6) |
| 5 | Claude Opus 5 (Max) | 0.22 | #16 (Claude Opus 5) |
| 6 | Claude Fable 5 (Max) | 0.21 | #20 (Claude Fable 5) |
| 7 | DeepSeek V4.1 Flash (1017) | 0.21 | #27 |
| 8 | Kimi K3 | 0.21 | #35 |
| 9 | Claude Fable 5 | 0.21 | #20 |
| 10 | GLM-5.3 | 0.18 | #44 |
| 11 | DeepSeek V4 Flash (0731) | 0.18 | #84 |
| 12 | MiniMax-M3 | 0.17 | #151 |
| 13 | Qwen 3.8 Max (xHigh) | 0.16 | #34 (Qwen 3.8 Max) |
| 14 | GPT-5.6 Sol (xHigh) | 0.15 | #22 (GPT-5.6 Sol) |
| 15 | GPT-5.5 (xHigh) | 0.14 | #33 (GPT-5.5) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kernelbench-hard · How It Works · Data refreshed daily, snapshot 2026-11-05.