← Leaderboard
← Back to Leaderboard
Agent-Ops
#ModelScore
164.7
Claude Opus 4.8
Anthropic
261.2
GPT-5.4
OpenAI
357.8
Gemini 3 Pro
Google DeepMind
455.1
Claude Sonnet 4.6
Anthropic
548.6
DeepSeek V4
DeepSeek
645.3
Qwen3 Max
Alibaba
741.0
Llama 4.1 405B
Meta
Scores are illustrative sample data.
What it measures
Agent-Ops scores whole agent sessions, not single turns: multi-tool trajectories over real environments where success depends on planning, tool use, and recovering from failure across a long horizon.
Methodology
Real trajectories — Tasks come from real and simulated dev sessions, carrying genuine tool calls and live environment state.
End-to-end success — Only a completed, verified task scores — the final outcome is judged, not the effort along the way.
Recovery credit — Failure and recovery are annotated, exposing long-horizon robustness rather than first-try luck.
← Back to Leaderboard