HarvestBench - Neutral Briefing: leaderboard
The same tractor game with the morality criterion removed from the briefing; the number is the share of animal encounters where the agent drove over the animal rather than swerving, which usually costs fuel. It shows what the agent does when nothing in its instructions mentions ethics.
Metric: Animal continue rate (%; share of animal encounters where the agent drove over the animal instead of paying fuel to swerve, lower is better). Source: arxiv.org. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Medium) | 84.6 | #30 (GPT-5.6 Sol) |
| 2 | GPT-5 Mini (Medium) | 91 | #348 (GPT-5 Mini) |
| 3 | Mistral Small 3.2 | 92.2 | #1591 |
| 4 | Claude Haiku 4.5 (Thinking) | 98.4 | #568 (Claude Haiku 4.5) |
| 5 | GPT-4o Mini | 98.8 | #1623 |
| 6 | Gemini 2.5 Flash (Thinking) | 100 | #453 (Gemini 2.5 Flash) |
| 7 | DeepSeek V3.1 (Thinking) | 100 | #497 (DeepSeek V3.1) |
| 8 | GPT-5.6 Terra (Medium) | 100 | #62 (GPT-5.6 Terra) |
Interactive version: theaggregate.ai/benchmark?slug=harvestbench-neutral-briefing · How It Works · Data refreshed daily, snapshot 2026-11-13.