EnvSimBench: leaderboard
Scalable AI agents training relies on interactive environments that faithfully simulate the consequences of agent actions. Manually crafted environments are expensive to build, brittle to extend, and fundamentally limited in diversity. A promising direction is to replace manually crafted environments with LLM-simulated counterparts. However, this paradigm hinges on an unexamined core assumption: L
Metric: CM Overall (self-reported). Source: benchmarklist.com. Saturation forecast: Around July 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 42.3 | #146 |
| 2 | Gemini 3.1 Pro (Preview) | 42 | #58 |
| 3 | GPT-5.4 | 42 | #79 |
| 4 | MiniMax-M2.7 | 41.8 | #258 |
| 5 | GLM-5 | 41 | #141 |
| 6 | Claude Sonnet 4.6 | 37.8 | #90 |
| 7 | DeepSeek V3.2 | 32.5 | #205 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=envsimbench · How It Works · Data refreshed daily, snapshot 2026-10-20.