HAS-Bench - Retail: leaderboard
HAS-Bench (2026) tests LLM agents that work with simulated human participants who hold task information, constraints or authority. This board is Pass@1 in percent on the 94 retail tasks (adapted from tau2-Bench) under Equal Partnership (A3), where the agent may ask a GPT-4.1-simulated human for clarification, feedback and control; a task passes only when every task dimension passes, and higher is better.
Metric: Pass@1 (%; share of the 94 retail tasks in which every task dimension passes) under Equal Partnership (A3) human participation: the agent may ask a GPT-4.1-simulated human for clarification, feedback and control, temperature 0, one rollout per task; GPT-4.1 judge; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 44.7 | #400 |
| 2 | GPT-4.1 | 42.6 | #495 |
| 3 | GPT-4.1 Mini | 39.4 | #779 |
| 4 | DeepSeek V3 (0324) | 39.4 | #709 |
| 5 | Llama 3.1 8B | 0 | #4861 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=has-bench-retail · How It Works · Data refreshed daily, snapshot 2026-11-13.