Atlas

Benchmarks

← All benchmarks

LLM Position Bias First-Shown Pick Rate

Chat & Writing · 2026-04-21

First-Shown Pick Rate is the share of decisive prompt responses where the model chose the version displayed first. The unbiased target is 50%; values above or below 50% indicate first- or second-position preference.

Top models (closest to 50.0 is better)

ModelScore
Trinity Large Thinking48.9
Doubao Seed 2.0 Pro48.1
Gemini 3.1 Flash-Lite Preview52.5
DeepSeek-V3.253.8
MiMo-V2-Pro54.5
Qwen3.5 122B A10B56.7
MiniMax M2.758.2
Claude Fable 558.9
GLM-5.159.8
Opus 4.860.7
Gemini 3.5 Flash62.1
Qwen3.6 Plus (2026-04-02)64.3
Opus 4.664.4
Sonnet 4.665.1
Qwen3.7-Max65.3
MiniMax M365.3
Opus 4.765.6
Grok 4.2065.8
Qwen3.5 397B A17B65.8
Gemini 3.1 Pro Preview66.0
ERNIE 5.066.1
DeepSeek-V4-Pro66.7
Mistral Medium 3.167.2
GPT-5.569.9
Kimi K2.672.5
Mistral Large 3 675B Instruct 251227.4
GPT-5.472.7
Gemma 4 31B IT74.6
GPT-5.4 Mini75.7
Kimi K2.575.8
Mistral Medium 3.582.8
Loading Atlas data…