Public benchmark radarTransparent evaluation
See how frontier agents perform on real work.
A small, readable view of public evaluations, with source links to the original leaderboards.
Long-horizon engineeringDeepSWE v1.1
Original repository tasks designed to measure how coding agents handle real, multi-file engineering work.
74%top pass@1
113 tasks91 repositories5 languages
modelpass@1avg cost
Claude Opus 574%10.06
GPT-5.6 Sol73%5.49
GLM 5.366.9%3.39
Kimi K369%3.95
DeepSeek V4 Pro63%1.42
