DeepSeek DSBench - Hard: leaderboard
DeepSeek's internal DSBench-Hard, a test set of difficult coding-agent problems. DeepSeek-V4-Pro-0813 model card table plus the DeepSeek-V4-Flash-Vision-Exp card, whose shared reference cells match; score in percent, higher is better. DeepSeek's card notes run its models' agent benchmarks in the minimal mode of DeepSeek Harness at max reasoning effort; other models' setup is not stated.
Metric: Score (%). Source: huggingface.co. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 71.7 | #39 |
| 2 | Claude Fable 5 | 68.3 | #20 |
| 3 | DeepSeek V4 Pro (0813) | 67.2 | #58 |
| 4 | DeepSeek-V4-Flash-Vision-Exp | 63.6 | #88 |
| 5 | Kimi K3 | 63 | #35 |
| 6 | DeepSeek V4 Flash (0731) | 59.6 | #84 |
| 7 | GLM-5.2 | 54.5 | #80 |
| 8 | DeepSeek V4 Pro Preview | 31.1 | #150 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=deepseek-dsbench-hard · How It Works · Data refreshed daily, snapshot 2026-11-05.