MLCR-AA Accuracy (October 2026 methodology)
Professional Work
Accuracy under Artificial Analysis' October 2026 MLCR-AA methodology: the share of responses that a majority of the three judges find free of any claim contradicting the source medical record, including claims outside the ground truth, averaged over three runs of 60 questions. It is a second reading of the run behind the MLCR-AA pass rate and stays diagnostic.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6.1 Terra | 98.3 |
| GPT-6.1 Sol | 96.7 |
| GPT-6 Astra | 95.6 |
| GPT-6 Sol | 95.0 |
| Kimi K3.1 Pro | 93.9 |
| GPT-6 Luna | 88.9 |
| GPT-5.5 | 87.8 |
| GPT-5.5 Instant (2026-06-25 hosted snapshot) | 81.1 |
| GPT-5.6 Terra | 78.3 |
| GPT-5.6 Sol | 76.7 |
| Qwen3.8 2.4T A95B | 76.1 |
| Qwen3.8-Max | 75.6 |
| DeepSeek V4.1 Flash 1017 | 67.8 |
| Qwen3.8 Max (0902) | 67.8 |
| Grok 4.7 | 67.2 |
| Muse Wonder | 65.6 |
| GPT-5.4 Nano | 65.0 |
| Beam 1.0 501B A23B | 64.4 |
| GPT-5.4 Mini | 64.4 |
| GPT-5.6 Luna | 63.9 |
| DeepSeek V4.1 Flash | 61.1 |
| Claude Sonnet 5.5 | 60.9 |
| Claude Haiku 5.5 | 58.9 |
| Muse Spark 1.3 | 58.9 |
| Muse Wonder Full | 58.9 |
| Qwen3.8 27B | 57.8 |
| Grok 4.6 | 55.6 |
| Qwen3.8-Flash-Next | 54.4 |
| Step 5 Preview | 49.4 |
| Grok 4.5 | 47.2 |
| K2 Horizon 375B A23B | 47.2 |
| Gemini 2.5 Pro | 46.7 |
| MiniMax M3.1 Flash | 46.7 |
| Kimi K3 | 39.4 |
| Opus 4.8 | 37.8 |
| Gemini 3.1 Pro Preview | 37.8 |
| Ling 3.1 Flash 1019 | 37.8 |
| Ling 3.1 Flash | 37.8 |
| Claude Opus 5.5 | 37.2 |
| Muse Spark 1.1 | 36.7 |