Humanity's Last Exam RMS Calibration Error (CAIS)
General QA · 2025-01-24
RMS calibration error on Humanity's Last Exam measures how closely a model's stated confidence matches its observed accuracy across expert-written academic questions. Lower calibration error indicates better confidence calibration.
Top models (lower is better)
| Model | Score |
|---|---|
| Muse Spark 1.1 | 25.3 |
| Claude Fable 5 | 25.5 |
| Opus 4.8 | 26.4 |
| Opus 4.7 | 28.1 |
| Grok 4.5 | 30.8 |
| GPT-5.4 | 40.9 |
| GPT-5.5 | 41.8 |
| GLM-5.2 | 42.0 |
| Grok 4.3 | 42.5 |
| Sonnet 4.6 | 43.1 |
| GPT-5.2 | 43.7 |
| Kimi K3 | 44.7 |
| Grok 4.20 | 45.8 |
| Opus 4.6 | 46.0 |
| GPT-5.6 Sol | 46.7 |
| GPT-5.4 Nano | 47.1 |
| GPT-5.1 | 47.8 |
| GPT-5 | 50.0 |
| Gemini 3.1 Pro Preview | 50.3 |
| GPT-5.6 Terra | 50.7 |
| Gemini 3.5 Flash | 53.5 |
| Gemini 3 Flash Preview | 55.1 |
| GPT-5.4 Mini | 55.2 |
| GPT-5.6 Luna | 55.7 |
| Grok 4 | 56.4 |
| Opus 4.5 | 56.9 |
| DeepSeek-V4-Pro | 57.1 |
| Gemini 3 Pro Preview | 57.2 |
| Kimi K2.6 | 58.2 |
| GLM-5.1 | 58.5 |
| DeepSeek-V3.2 | 58.8 |
| Haiku 4.5 | 60.5 |
| Sonnet 4.5 | 64.1 |
| GPT-5 Mini | 65.0 |
| Kimi K2 Thinking | 67.8 |
| Grok 4 Fast | 68.7 |
| Gemini 2.5 Pro Preview 03-25 | 71.0 |
| Gemini 2.5 Pro | 72.0 |
| Grok 4.1 Fast | 72.1 |
| Gemini 3.1 Flash-Lite Preview | 72.8 |