How well a model works through a task with tools rather than answering in one turn. Two questions live here and they are not the same one: which model to drive, and which harness to drive it with.
Each panel is one source's own ranking, shown as it publishes it. Nothing on this page is averaged across panels — where two disagree, both are right about what they measured.
The model itself, whatever it was driven by. Artificial Analysis grades completed tasks; Arena observes real sessions and scores tool reliability, task completion and steerability over its own traffic.
Agentic Index fetched 2026-08-22
Claude Opus 5 59.2 
GLM-5.3 59.1 
Grok 4.6 58.7 
Qwen3.8 Max 58.4 
GPT-5.6 Sol 57.8 Qwen3.8 2.4T A95B 57.1 Claude Fable 5 (Opus 4.8 fallback) 56.6 
Kimi K3 54.3 Qwen3.8 27B 50.9 GPT-5.6 Terra 50.2 Artificial Analysis Coding Agent Index fetched 2026-08-20
Claude Opus 5 0.667 GPT-5.6 Sol 0.666 Grok 4.5 0.644 GPT-5.6 Terra 0.623 GPT-5.5 0.615 Kimi K3 0.613 Claude Opus 4.8 0.606 
Muse Spark 1.2 0.605 Qwen3.8 Max 0.598 Overall as of 2026-08-19
Claude Opus 5 0.125 Kimi K3 0.104 GPT-5.6 Sol 0.097 Claude Opus 4.8 0.095 GPT-5.5 0.085 Claude Opus 4.7 0.081 Claude Sonnet 5 0.066 Claude Opus 4.6 0.066 The same rows read the other way round: an agent is a program, and the model is what it drives. This has one source, because it is the only board found that names both halves of the pair.
Artificial Analysis Coding Agent Index by harness · fetched 2026-08-20