AgentGym2: leaderboard

Metric: Task success (%; Avg@3, all 437 tasks, task-count weighted over the three scenarios; 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-546.15
2Claude Sonnet 4.5 (Thinking)37.17
3DeepSeek V3.1 (Non-reasoning)21.66
4DeepSeek V3.2 Exp (Non-reasoning)20.98
5Kimi K220.9
6Gemini 2.5 Pro20.67
7GLM-4.620.52
8Qwen 3 235B A22B 2507 Instruct14.04
9Qwen 3 235B A22B 2507 (Thinking)9.76
10Qwen 3 32B (Thinking)7.4
11Qwen 3 32B (Non-reasoning)6.55
12Qwen 3 8B (Thinking)5.49
13Qwen 3 8B (Non-reasoning)4.12

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