ClawArena (Claude Code): leaderboard

ClawArena (2026) tests agents as persistent assistants in evolving information environments: 12 multi-turn scenarios with conflicting sources, staged updates and implicit preferences, checked by multiple-choice and executable shell questions, here with the agent run inside the Claude Code harness. Composite reliability score from 0 to 100, the mean of task completion and robustness; higher is better.

Metric: Composite Reliability Score (CRS, 0-100) on ClawArena (12 multi-turn scenarios, 337 evaluation rounds with staged updates; multi-choice and shell executable-check questions): the mean of the task completion rate and Robustness (success cohesion times failure dispersion), macro-averaged over scenarios; agent run in the Claude Code framework; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.766.31#85
2Claude Sonnet 4.662.16#173
3Claude Haiku 4.560.93#583
4Kimi K2.559.75#266
5GPT-5.147.57#240
6GPT-5.544.5#54

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=clawarena-claude-code · How It Works · Data refreshed daily, snapshot 2026-11-21.