AgentGym2: leaderboard
AgentGym2 (2026) tests LLM agents on 437 real-world end-to-end tasks where tools must be discovered by exploration and inputs are noisy or underspecified. Success is judged by an LLM against short reference answers and averaged over three runs; higher is better, in percent.
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 February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 46.15 | #167 |
| 2 | Claude Sonnet 4.5 (Thinking) | 37.17 | #250 (Claude Sonnet 4.5) |
| 3 | DeepSeek V3.1 (Non-reasoning) | 21.66 | #497 (DeepSeek V3.1) |
| 4 | DeepSeek V3.2 Exp (Non-reasoning) | 20.98 | #438 (DeepSeek V3.2 Exp) |
| 5 | Kimi K2 | 20.9 | #439 |
| 6 | Gemini 2.5 Pro | 20.67 | #266 |
| 7 | GLM-4.6 | 20.52 | #486 |
| 8 | Qwen 3 235B A22B 2507 Instruct | 14.04 | #619 |
| 9 | Qwen 3 235B A22B 2507 (Thinking) | 9.76 | #463 |
| 10 | Qwen 3 32B (Thinking) | 7.4 | #960 (Qwen 3 32B) |
| 11 | Qwen 3 32B (Non-reasoning) | 6.55 | #960 (Qwen 3 32B) |
| 12 | Qwen 3 8B (Thinking) | 5.49 | #1830 (Qwen 3 8B) |
| 13 | Qwen 3 8B (Non-reasoning) | 4.12 | #1830 (Qwen 3 8B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agentgym2 · How It Works · Data refreshed daily, snapshot 2026-11-13.