CAB - Race: leaderboard
The model answers realistic open-ended user questions asked twice, differing only in the stated race; a GPT-5-mini judge scores how strongly the two answers differ in ways the question does not justify and the model does not acknowledge, averaged per model (0-3.2). Unlike BBQ, there is no multiple-choice answer key: the score is about unequal treatment in free-text help.
Metric: Mean counterfactual bias fitness (0-3.2, lower is better). Source: arxiv.org. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 1.47 | #400 |
| 2 | GPT-OSS-120B (Low) | 1.65 | #700 (GPT-OSS-120B) |
| 3 | GPT-5 Chat | 1.66 | #444 |
| 4 | DeepSeek V3.1 (Non-reasoning) | 1.73 | #497 (DeepSeek V3.1) |
| 5 | Grok 4 | 1.83 | #307 |
| 6 | Kimi K2 | 1.83 | #439 |
| 7 | GLM-4.5 (Non-reasoning) | 1.83 | #526 (GLM-4.5) |
| 8 | Qwen 3 235B A22B 2507 Instruct | 1.94 | #619 |
Interactive version: theaggregate.ai/benchmark?slug=cab-race · How It Works · Data refreshed daily, snapshot 2026-11-13.