DeepSeek V4 Pro (0813) across 5 agent harnesses

Success rate

Higher is better
100%
80%
60%
40%
20%
0%
70%
Pi
Agent
67%
DeepSeek
Harness
63%
Claude
Code
63%
OpenCode
60%
Hermes
Agent

Cost per success

Lower is better
$0.080
$0.064
$0.048
$0.032
$0.016
$0.000
$0.031
Pi
Agent
$0.028
DeepSeek
Harness
$0.074
Claude
Code
$0.032
OpenCode
$0.037
Hermes
Agent

Median time per task

Lower is better
400s
320s
240s
160s
80s
0s
363s
Pi
Agent
252s
DeepSeek
Harness
182s
Claude
Code
281s
OpenCode
274s
Hermes
Agent
ComposioBENCH
Source: Composio agentic eval