EnterpriseBench Sandbox (LangChain, ReAct): leaderboard
EnterpriseBench (Vishwakarma et al., Fujitsu Research, 2025; not the Surge AI board of the same name) places an LLM agent in a simulated company behind role-based access control and gives it 500 search, create-read-update-delete and unanswerable tasks across five departments. This board runs the agent built in LangChain with two few-shot examples per call and ReAct-style interleaved reasoning and tool calls. GPT-4 grades each final execution with the Prometheus-2 rubric, and the score is task accuracy from 0 to 1, higher is better.
Metric: Judged task accuracy (0-1; 500 enterprise tasks, final execution graded with the Prometheus-2 rubric by GPT-4; agent built in LangChain with two few-shot examples per call, ReAct-style planning). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 Mini | 0.35 | #761 |
| 2 | GPT-4o | 0.32 | #741 |
| 3 | Claude 3.5 Sonnet (20240620) | 0.28 | #684 |
| 4 | Llama 3.3 70B Instruct | 0.21 | #1353 |
| 5 | Llama 3.1 8B Instruct | 0.14 | #4277 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=enterprisebench-sandbox-langchain-react · How It Works · Data refreshed daily, snapshot 2026-11-12.