HarvestBench: leaderboard
An agent drives a tractor through a pasture to bring in a corn harvest and, at each animal in its path, chooses to drive over it at no cost or to swerve, which usually costs fuel (about a fifth of swerves cost none); the number is the share of animal encounters where it drove on, under a briefing that scores 'moral conduct' without naming animals. Unlike TAC (Travel Agent Compassion), which checks whether an agent raises an unmentioned harm, the animal and the cost of avoiding it are shown to the agent.
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. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Terra (Medium) | 0.4 | #62 (GPT-5.6 Terra) |
| 2 | GPT-5.6 Sol (Medium) | 0.9 | #30 (GPT-5.6 Sol) |
| 3 | DeepSeek V3.1 (Thinking) | 2.4 | #497 (DeepSeek V3.1) |
| 4 | Claude Haiku 4.5 (Thinking) | 4.5 | #568 (Claude Haiku 4.5) |
| 5 | GPT-5 Mini (Medium) | 5.4 | #348 (GPT-5 Mini) |
| 6 | Claude Sonnet 5 (Medium) | 17.8 | #111 (Claude Sonnet 5) |
| 7 | Gemini 2.5 Flash (Thinking) | 38.7 | #453 (Gemini 2.5 Flash) |
| 8 | Mistral Small 3.2 | 88.8 | #1591 |
| 9 | GPT-4o Mini | 98.8 | #1623 |
Interactive version: theaggregate.ai/benchmark?slug=harvestbench · How It Works · Data refreshed daily, snapshot 2026-11-13.