ClinSQL: leaderboard
ClinSQL (Yale University and University of Washington, 2026) tests clinical text-to-SQL on MIMIC-IV v3.1 with 633 expert-written tasks that require patient-similarity cohort construction, temporal windows and clinical coding, scored here on the test split. Models write one BigQuery query under chain-of-thought prompting with up to two self-refinement rounds on execution errors, and a GPT-5 judge scores the executed result table against an expert rubric tree with critical-first aggregation. Execution Score is the mean in percent, higher is better.
Metric: Execution Score (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini (2025-08-07) | 74.67 | #341 |
| 2 | Gemini 2.5 Pro | 73.73 | #262 |
| 3 | DeepSeek R1 0528 | 69.15 | #425 |
| 4 | GPT-4.1 | 67.79 | #492 |
| 5 | Gemini 2.5 Flash | 65.01 | #452 |
| 6 | DeepSeek V3.1 | 60.71 | #500 |
| 7 | O4 Mini (2025-04-16) | 59.22 | #362 |
| 8 | Qwen 3 235B A22B 2507 Instruct | 58.63 | #619 |
| 9 | Grok 4 Fast (Reasoning) | 58.46 | #438 (Grok 4 Fast) |
| 10 | Qwen 3 Coder 480B A35B Instruct | 58.18 | #641 |
| 11 | GPT-5 Nano | 52.03 | #888 |
| 12 | Qwen 3 Next 80B A3B Instruct | 49.26 | #763 |
| 13 | Qwen 3 235B A22B 2507 FP8 (Thinking) | 48.54 | #431 |
| 14 | Llama 4 Maverick Instruct | 47.49 | #1125 |
| 15 | Mistral Medium 3 | 45.1 | #1079 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=clinsql · How It Works · Data refreshed daily, snapshot 2026-11-12.