GenderBench: leaderboard
The mean of GenderBench's 19 gender-bias probe metrics, each normalized by the publisher to 0 (least biased) to 1 (most biased), for chat models asked to make decisions, answer questions and write stories. Unlike BBQ alone, it mixes outcome disparity, stereotypical reasoning and representation probes.
Metric: Mean normalized bias over 19 probe metrics (0-1, lower is better). Source: genderbench.readthedocs.io. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.0 Flash | 0.13 | #704 |
| 2 | Gemini 2.0 Flash Lite | 0.13 | #1032 |
| 3 | Mistral Small 3 | 0.14 | #1777 |
| 4 | Llama 3.1 8B Instruct | 0.14 | #4246 |
| 5 | GPT-4o | 0.15 | #741 |
| 6 | Mistral 7B Instruct (v0.3) | 0.15 | #6653 |
| 7 | Gemma 2 9B (IT) | 0.15 | #2262 |
| 8 | Phi-4 | 0.15 | #1913 |
| 9 | GPT-4o Mini | 0.15 | #1623 |
| 10 | Gemma 2 27B (IT) | 0.16 | #1782 |
| 11 | Claude 3.5 Haiku | 0.16 | #1553 |
| 12 | Llama 3.3 70B Instruct | 0.17 | #1357 |
Interactive version: theaggregate.ai/benchmark?slug=genderbench · How It Works · Data refreshed daily, snapshot 2026-11-13.