ERPBench (ERPNext) - Single-Record Tasks: leaderboard
ERPBench (Bhagtani et al., Accenture, 2026) tests screenshot-only computer-use agents on a live, self-hosted ERPNext enterprise resource planning system and grades each task by the values left in the ERP database, not by what appears on screen. This board covers the 20 single-record tasks (field edits, selections and simple record creation), five runs each; success rate in percent, higher is better.
Metric: Success rate (%; tier-1 single-record tasks on a self-hosted ERPNext instance: 20 tasks (text edits, selects and toggles, simple record creation), five runs each, 100 runs; the agent sees only 1280x720 screenshots and acts with pixel-coordinate mouse and keyboard actions in the authors' harness with auto-approval; a run succeeds when every target field read back from the ERPNext database matches the ground truth). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 94 | #158 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=erpbench-erpnext-single-record-tasks · How It Works · Data refreshed daily, snapshot 2026-11-12.