EnvTrustBench: leaderboard
Evidence grounding of LLM coding agents (Sheng, Wang and Zhou, 2026): 55 generated operational cases in 11 scenarios plant stale, wrong or malicious claims in files, tool outputs and artifacts, and an oracle checks whether the agent acted on them instead of the true environment state. Environmental misgrounding rate in percent over five runs per case, per agent scaffold and model; lower is better.
Metric: Environmental misgrounding rate (%; share of accepted pass-or-fail runs that complete the case's false path, mean over 11 operational scenarios of 5 generated cases x 5 runs each). Source: arxiv.org. Saturation forecast: Around 2030. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 55.3 | #158 |
| 2 | GLM-5.1 | 72 | #175 |
| 3 | Gemini 3.1 Pro (Preview) | 85.8 | #91 |
| 4 | Qwen 3.6 Plus | 88.4 | #217 |
| 5 | GPT-5.5 | 88.7 | #48 |
| 6 | DeepSeek V4 Pro | 95.6 | #181 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=envtrustbench · How It Works · Data refreshed daily, snapshot 2026-11-12.