DSBench-Hard

DSBench-Hard - internal test set of difficult coding-agent problems for evaluating agentic coding on challenging software engineering tasks.

Category

Agentic Coding

Max Score

100.0

Score Type

percent

Active

Yes

Model Scores

Scores for DSBench-Hard

Model Score Date Verified
DeepSeek-V4-Flash-0731
73.64%
31.07.2026 Verified
DeepSeek-V4-Flash
25.80
13.08.2026 Verified
GLM-5.2
62.53%
31.07.2026 Verified
DeepSeek-V4-Pro
11.55%
13.08.2026 Verified
Opus-4.8
100.00%
31.07.2026 Verified
DeepSeek-V4-Pro-0813
90.20%
13.08.2026 Verified
Kimi K3
81.05%
13.08.2026 Verified
Fable-5
92.59%
13.08.2026 Verified
DeepSeek-V4-Flash-Vision-Exp
82.35%
08.09.2026 Verified