EnvSimBench - Difficult: leaderboard

The Difficult tier of EnvSimBench (2026): 50 tool calls whose correct outcome changes 7 to 12 changed state fields of the environment state. The model sees the state, the action and its code and must predict the resulting state; the score is Config Match, the percent of samples reproduced exactly, with non-thinking mode throughout; higher is better.

Metric: Config Match (%): share of the 50 Difficult state-changing samples (7 to 12 changed state fields) whose predicted state changes, applied to the before-state, reproduce the executor after-state exactly; MDP prompt with the state, the action and its code, non-thinking mode; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GLM-5Non-reasoning28#276
2GPT-5.4Non-reasoning26#138
3Qwen 3.5 397B A17BNon-reasoning24#270
4MiniMax-M2.722#539
5Qwen 3.5 9BNon-reasoning20#838
6Gemini 3.1 Pro (Preview)18#102
7Claude Sonnet 4.616#179
8Qwen 3.5 4BNon-reasoning6#1270
9DeepSeek V3.2Non-reasoning4#425
10Qwen 3.5 27BNon-reasoning4#380

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=envsimbench-difficult · How It Works · Data refreshed daily, snapshot 2026-11-22.