Lumen-bench: leaderboard
Tool-calling models face 108 ethical scenarios under social pressure in five languages with 21 generic tools; the score is the share of valid trials, out of 6,480 scheduled per model, in which the model made a harmful tool call, whatever it said in text. Unlike AgentHarm it scores actions on everyday ethical dilemmas (privacy, manipulation, whistleblowing) and compares languages.
Metric: Harmful tool-action rate (%, lower is better). Source: arxiv.org. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0 | #94 |
| 2 | GPT-5.4 | 0 | #123 |
| 3 | Claude Haiku 4.5 | 0 | #568 |
| 4 | GPT-5 Mini (2025-08-07) | 0.05 | #343 |
| 5 | MiniMax-M2.7 | 0.08 | #518 |
| 6 | Gemma 4 31B (IT) | 0.23 | #342 |
| 7 | Qwen 3.5 27B | 0.25 | #359 |
| 8 | DeepSeek V3.2 | 0.28 | #406 |
| 9 | Qwen 3.5 397B A17B | 0.34 | #246 |
| 10 | Gemma 4 26B A4B (IT) | 0.39 | #549 |
| 11 | GPT-OSS-120B | 0.56 | #700 |
| 12 | GPT-OSS-20B | 0.88 | #1147 |
| 13 | Qwen 3.5 35B A3B | 0.88 | #482 |
| 14 | GPT-4.1 | 1.5 | #495 |
| 15 | Gemini 3.1 Flash Lite (Preview) | 2.31 | #418 |
Interactive version: theaggregate.ai/benchmark?slug=lumen-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.