ToolMaze - Perturbation Recovery Rate: leaderboard

ToolMaze (2026) tests tool-using LLM agents on 400 DAG-structured tasks over 270 tools, each also run with injected tool failures that are explicit or implicit and transient or permanent. This board is the percentage of failure-exposed runs in which the agent retries, switches tool path or aborts correctly, under the standard tool-use prompt; higher is better.

Metric: Perturbation recovery rate (%): share of runs exposed to an injected failure in which the agent retries a transient fault, switches to an alternative tool path or aborts an unsolvable task, averaged over the four failure modes; 400 DAG-structured tool-use tasks (four topology levels) over 270 tools, each also run under four injected tool-failure modes (explicit or implicit, transient or permanent); sandboxed, 25 agent steps, temperature 1, the standard tool-use prompt; higher is better. Source: arxiv.org. Saturation forecast: Around September 2028. 9 models tracked. Updated 30 Sep 2026.

Top models

#ModelScoreOverall rank
1GPT-5.537.31#52
2DeepSeek V4 Pro36.71#191
3Gemini 3.1 Pro (Preview)31.48#93
4GLM-5.130.43#186
5Qwen 3.5 397B A17B29.58#251
6Qwen 3.5 35B A3B29.48#486
7Qwen 3.6 27B28.75#344
8Claude Sonnet 4.624.88#165
9MiniMax-M2.723.34#523

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=toolmaze-perturbation-recovery-rate · How It Works · Data refreshed daily, snapshot 2026-11-16.