Bedside Bench
Professional Work
Doximity's physician-validated clinical benchmark in its finalized Slack-500 form: 500 synthetic or de-identified cases in ten 50-case suites covering clinical calculators and calculations, drug safety, guideline adherence, health equity, landmark trials, medical hallucination, and diagnostic and therapeutic safety. Each answer is judged by an LLM against physician-authored signed point rubrics (weighted F1 over appropriate and inappropriate actions for the two management-plan suites), and the score is the unweighted macro-average of the ten suite means. Fireworks runs it for its Specialized Intelligence Index with GPT-5.6 Sol as judge.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 94.6 |
| Muse Spark Open | 94.3 |
| GPT-6 Astra | 93.4 |
| Gemini 4 Argon | 92.7 |
| GPT-5.6 Sol | 91.1 |
| Kimi K3 | 89.7 |
| Ember-1 | 89.1 |
| GPT-6 Sol | 88.1 |
| GPT-6 Luna | 87.2 |
| GLM-5.3 | 86.9 |
| Gemini 3.8 Flash | 86.5 |
| DeepSeek V4.1 Flash 1017 | 85.7 |
| Grok 4.6 | 83.3 |
| DeepSeek V4.1 Flash | 82.9 |
| DeepSeek V4 Pro 0813 | 80.6 |