SweetBench - COGI: leaderboard
SweetBench Table 3:212 simulated 72-hour tasks, Dana-i pump, insulin aspart,real-timeCGM,I 2 observations; APC 60-minute decisions and high reasoning effort. Scores average completed episodes equally from 864 true-glucose samples; completion-conditioned control scores, not clinical outcomes or overall task success. Model backends have the distinct harnesses recorded in Table 11. Metric: Control score (0–100).
Metric: Control score (0–100). Source: arxiv.org. Saturation forecast: Around October 2028. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (High) | 73.54 | #3 (Claude Opus 5.5) |
| 2 | GPT-6.1 Sol (High) | 70.58 | #7 (GPT-6.1 Sol) |
| 3 | DeepSeek V4.1 Flash (High) | 68.61 | #37 (DeepSeek V4.1 Flash) |
| 4 | Kimi K3.1 Pro (High) | 67.15 | #17 (Kimi K3.1 Pro) |
| 5 | GLM-5.3 Flash (High) | 66.25 | #77 (GLM-5.3 Flash) |
No result here: #6 GPT-6 Astra, #9 Gemini 4 Argon, #10 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=sweetbench-cogi · How It Works · Data refreshed daily, snapshot 2026-11-05.