Best AI models for base-rate bias
Ranked by accuracy on the meo “base-rate bias” domain — one of eleven un-leaked, contamination-resistant domains. Per-domain strength reveals where a model genuinely reasons versus where it pattern-matches; the leaders differ sharply by domain.
Ranking 22 models across the meo-tested roster · as of 2026-06-08.
| # | Model | Lab | accuracy |
|---|---|---|---|
| 1 | Anthropic: Claude Opus 4.8 | anthropic | 10% |
| 2 | Owl Alpha | openrouter | 10% |
| 3 | Z.ai: GLM 5.1 | z-ai | 10% |
| 4 | Google: Gemini 3.1 Flash Lite | 10% | |
| 5 | OpenAI: GPT-5.5 | openai | 7% |
| 6 | inclusionAI: Ring-2.6-1T | inclusionai | 7% |
| 7 | Xiaomi: MiMo-V2.5 | xiaomi | 7% |
| 8 | MiniMax: MiniMax M3 | minimax | 7% |
| 9 | Perceptron: Perceptron Mk1 | perceptron | 7% |
| 10 | Tencent: Hy3 preview | tencent | 7% |
| 11 | Arcee AI: Trinity Large Thinking | arcee-ai | 7% |
| 12 | Qwen: Qwen3.7 Max | qwen | 5% |
| 13 | DeepSeek: DeepSeek V4 Flash | deepseek | 5% |
| 14 | DeepSeek: DeepSeek V4 Pro | deepseek | 5% |
| 15 | Qwen: Qwen3.7 Plus | qwen | 5% |
| 16 | Xiaomi: MiMo-V2.5-Pro | xiaomi | 5% |
| 17 | Google: Gemini 3.1 Pro Preview | 2% | |
| 18 | Google: Gemini 3.5 Flash | 2% | |
| 19 | MoonshotAI: Kimi K2.6 | moonshotai | 2% |
| 20 | NVIDIA: Nemotron 3 Ultra | nvidia | 2% |
| 21 | StepFun: Step 3.7 Flash | stepfun | 2% |
| 22 | xAI: Grok 4.3 | x-ai | 0% |