CARE-Bench: leaderboard
Unprompted four-label current-action macro-F1 over the combined 269-round held-out test set. Responses are capped at 300 tokens and scored by a fixed mapper; this is neither diagnostic accuracy nor the prompted protocol. Local healthcare models use 4-bit NF4 quantization.
Metric: Macro-F1 (%, unprompted). Source: arxiv.org. Saturation forecast: Around 2033. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 4 Scout | 50.4 | #656 |
| 2 | Claude Opus 4.7 | 48.7 | #49 |
| 3 | GPT-5.5 | 46.2 | #31 |
| 4 | Claude Haiku 4.5 | 44.5 | #280 |
| 5 | Mistral Large 3 | 43.2 | #401 |
| 6 | GPT-5.4 Mini | 43.1 | #212 |
| 7 | Qwen 3 32B | 42.2 | #443 |
| 8 | Gemini 1.0 Pro | 34.6 | #933 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=care-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.