MedScribe
Whether a model can write clinical SOAP notes — Subjective, Objective, Assessment, Plan — accurately enough to take the documentation burden off clinicians.
How it is run
Built with Protege: 100 rubrics score note quality against the encounter record. Vals also reports average output length, because note length and accuracy move together on this board.
What it cannot tell you
The leaders sit within about three points of each other near 90%, so the overall number separates less than it appears to. Most models do worst on the Plan section — orders, prescriptions, referrals and follow-ups — where documentation errors carry the most clinical and billing risk, which makes that the part still needing human review.
- Leader
- Claude Opus 5.591.43% · Vals AI's own list
- Top-ten spread
- 2.91%Leader minus tenth place, or last place in smaller fields.
- Median
- 80.68%of the 106 rows published
- No published result
- 10of 46 models on this site
Independent leaderboard · Vals AI
Top 106 of 106 runs by Vals AI, one harness.
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 5.5 | 91.43% |
| 2 | Claude Fable 5.1 | 91.29% |
| 3 | Claude Sonnet 5.5 | 91.10% |
| 4 | Claude Opus 5 | 90.98% |
| 5 | Muse Spark 1.2 | 90.06% |
| 6 | Grok 4.7 | 89.38% |
| 7 | GLM 5.3 Flash | 88.94% |
| 8 | Muse Spark 1 1 | 88.89% |
| 9 | GLM 5.3 | 88.81% |
| 10 | Claude Fable 5 | 88.52% |
| 11 | MiMo V2.6 Pro | 88.31% |
| 12 | Gpt 5.1 2025 11 13 | 88.09% |
| 13 | Kimi K3 | 87.96% |
| 14 | GPT-6 Astra | 87.91% |
| 15 | Gemini 4 Argon | 87.43% |
| 16 | MiniMax M3 | 87.25% |
| 17 | Grok 4.5 | 86.88% |
| 18 | GPT-5.5 | 86.87% |
| 19 | Claude Opus 4 6 | 86.74% |
| 20 | Grok 4.6 | 86.53% |
| 21 | GPT-6.1 Sol | 86.45% |
| 22 | Claude Opus 4 6 Thinking | 86.13% |
| 23 | Muse Spark | 85.90% |
| 24 | Claude Opus 4.8 | 85.75% |
| 25 | DeepSeek V4.1 Flash | 85.50% |
| 26 | Inkling | 85.41% |
| 27 | Claude Opus 4 5 20251101 Thinking | 85.32% |
| 28 | MiMo-V2.6-Flash | 85.28% |
| 29 | GPT-5.6 Sol | 85.23% |
| 29 | Claude Haiku 4 5 20251001 Thinking | 85.23% |
| 31 | Qwen3.8 Max | 84.95% |
| 32 | Claude Sonnet 4 5 20250929 | 84.52% |
| 33 | Gemini 3.8 Flash | 84.50% |
| 34 | GPT-5.6 Luna | 84.39% |
| 34 | Gpt 5.2 2025 12 11 | 84.39% |
| 36 | Inkling Small | 84.11% |
| 37 | Claude Sonnet 4 5 20250929 Thinking | 84.10% |
| 38 | Gemini 3.7 Flash | 83.94% |
| 39 | Qwen3.8 27b | 83.85% |
| 40 | Mimo V2.5 Pro | 83.73% |
| 41 | GPT-6 Luna | 83.71% |
| 42 | Gpt 5 2025 08 07 | 83.65% |
| 43 | Hunyuan Hy4 Preview | 83.60% |
| 44 | Glm 5.2 | 83.53% |
| 45 | Claude Opus 4 5 20251101 | 83.25% |
| 46 | Gemini 2.5 Flash Thinking | 82.98% |
| 47 | Claude Opus 4.7 | 82.95% |
| 48 | Gemini 2.5 Flash | 82.87% |
| 48 | GPT-5.6 Terra | 82.87% |
| 50 | GPT-6 Sol | 82.03% |
| 51 | Grok 4 Fast Reasoning | 81.63% |
| 52 | Ling 3.0 Flash 2607 | 80.90% |
| 53 | MiniMax M2.1 | 80.78% |
| 54 | Gpt 5 Mini 2025 08 07 | 80.58% |
| 55 | Deepseek V4 Flash 0731 | 80.36% |
| 56 | DeepSeek V4 Pro | 80.17% |
| 57 | MiniMax M2.7 | 79.87% |
| 58 | Grok 4 Fast Non Reasoning | 79.72% |
| 59 | Gemini 3.6 Flash | 79.66% |
| 60 | Qwen3.7 Max | 79.40% |
| 61 | Grok 4 1 Fast Reasoning | 78.73% |
| 62 | Gemini 2.5 Flash Preview 09 2025 Thinking | 78.50% |
| 63 | Grok 4 0709 | 78.15% |
| 63 | Kimi K2.6 | 78.15% |
| 65 | Gemini 2.5 Flash Preview 09 2025 | 77.95% |
| 66 | Gpt 5.4 2026 03 05 | 77.55% |
| 67 | Grok 4 1 Fast Non Reasoning | 77.46% |
| 68 | Qwen3 Vl Plus 2025 09 23 | 77.13% |
| 69 | Gpt 5.4 Nano 2026 03 17 | 77.09% |
| 70 | Qwen3.6 Plus | 76.96% |
| 71 | O3 2025 04 16 | 76.65% |
| 72 | Gemini 3.5 Flash | 76.57% |
| 73 | Kimi K2.5 Thinking | 76.44% |
| 74 | Gemini 3.1 Pro | 76.11% |
| 75 | Claude Sonnet 5 | 76.05% |
| 76 | Gemini 2.5 Flash Lite Preview 09 2025 | 75.82% |
| 77 | Ling 3.0 Flash Af Rc3 | 75.59% |
| 78 | Deepseek V4 Pro | 75.14% |
| 79 | Grok 4.3 | 74.40% |
| 80 | Claude Opus 4 1 20250805 Thinking | 73.90% |
| 81 | Gemini 2.5 Pro | 73.55% |
| 82 | Gpt 5 Nano 2025 08 07 | 72.86% |
| 83 | Gemini 2.5 Flash Lite | 72.83% |
| 84 | Qwen3 Max 2026 01 23 | 72.71% |
| 85 | Claude Sonnet 4 20250514 | 72.41% |
| 86 | Glm 5.1 | 72.27% |
| 87 | Mimo V2.5 | 72.15% |
| 88 | Gemini 3 Pro Preview | 72.04% |
| 89 | Claude Opus 4 1 20250805 | 71.75% |
| 90 | Gemini 3.5 Flash Lite | 70.89% |
| 91 | Qwen3.5 Flash | 70.62% |
| 92 | Gemini 3 Flash Preview | 69.92% |
| 93 | Claude Sonnet 4 20250514 Thinking | 69.35% |
| 94 | O4 Mini 2025 04 16 | 69.14% |
| 95 | Glm 4.7 | 68.63% |
| 96 | Mistral Medium 3.5 | 67.73% |
| 97 | Gemini 2.5 Flash Lite Preview 09 2025 Thinking | 66.88% |
| 98 | Laguna M.1 | 65.91% |
| 99 | Gemini 3.1 Flash Lite Preview | 63.90% |
| 100 | Grok 4.20 0309 Reasoning | 63.41% |
| 101 | Laguna Xs.2 | 61.43% |
| 102 | Command A Plus 05 2026 | 55.68% |
| 103 | Mercury 2.5 | 55.09% |
| 104 | Llama4 Maverick Instruct Basic | 54.22% |
| 105 | Llama 4 Scout 17B 16E Instruct | 50.59% |
| 106 | Nemotron Lightning 3p5 30b A3b | 4.27% |
No published result tracked
10 of the 46 models on this site have no published number on this benchmark, so their cells in the home table stay blank rather than reading as zero.
Read this with
Benchmarks in the same category, which usually disagree with this one on purpose.
Sources
Every page used for this benchmark, read on 17 October 2026.
- Benchmark site or the page that publishes itwww.withprotege.ai/
- Independent resultswww.vals.ai/benchmarks/medscribe