InteractiveBench Prevention (task pass)
Agents · 2026-10-25
InteractiveBench's Prevention category: the 20 Prevention tasks, in which the assistant must warn the user before a mistake happens without interrupting needlessly (for example reminding them that the fridge and a window were left open before they leave). Same simulator, Gemini 3.1 Pro Human agent and end-state check as the full benchmark; the score is the percentage of these 20 tasks in which every required end state holds.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 50.0 |
| Opus 4.8 | 35.0 |
| Claude Opus 5 | 30.0 |
| GPT-5.6 Luna | 30.0 |
| Muse Spark 1.1 | 30.0 |
| GPT-5.5 | 25.0 |
| GLM-5.3 Flash | 20.0 |
| Kimi K3 | 20.0 |
| GPT-5.6 Sol | 15.0 |
| Gemini 3.1 Pro Preview | 10.0 |