AgentGym2 - Complex Tool Usage: leaderboard

Metric: Task success (%; Avg@3 on the 182 complex tool usage tasks; mean over three runs of the share of tasks answered correctly (Avg@3), short string or numeric answers graded CORRECT by a Qwen3-235B-A22B-Instruct-2507 judge; ReAct agent with OpenAI tool schemas, tools to be discovered in a de-idealized environment with noisy or underspecified inputs, at most 100 interaction rounds, temperature 0.6, thinking mode as printed per model). Source: arxiv.org. Saturation forecast: Around December 2026. 15 models tracked.

Top models

#ModelScore
1GPT-548.72
2Claude Sonnet 4.5 (Thinking)42.38
3DeepSeek V3.1 (Non-reasoning)23.26
4GLM-4.622.16
5Gemini 2.5 Pro21.25
6Kimi K219.96
7DeepSeek V3.2 Exp (Non-reasoning)19.6
8Qwen 3 235B A22B 2507 Instruct12.27
9Qwen 3 235B A22B 2507 (Thinking)7.33
10Qwen 3 32B (Thinking)6.41
11Qwen 3 32B (Non-reasoning)6.21
12Qwen 3 8B (Thinking)4.58
13Qwen 3 8B (Non-reasoning)2.75

Interactive version: theaggregate.ai/benchmark?slug=agentgym2-complex-tool-usage · How It Works · Data refreshed daily, snapshot 2026-09-29.