SM Bench — Overfit
Safety · 2026-02-01
SM Bench Overfit tests whether surface-level risk cues and policy-shaped phrasing override benign intent, common-sense reasoning, or useful task completion. Scores use difficulty-weighted pass, partial, and fail judgments across 100 fixed prompts; this category receives 2× weight in the overall SM Bench score.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.5 Flash | 98.4 |
| Gemini 3.1 Flash-Lite Preview | 97.8 |
| Opus 4.6 | 95.6 |
| Gemini 3 Flash Preview | 94.0 |
| Grok 4.5 | 92.3 |
| Opus 4.7 | 91.8 |
| Gemma 4 31B IT | 91.8 |
| Sonnet 4.6 | 91.5 |
| MiMo-V2-Pro | 90.7 |
| Opus 4.5 | 88.5 |
| Opus 4.8 | 87.4 |
| MiMo-V2.5-Pro | 85.8 |
| Gemini 3.1 Pro Preview | 84.2 |
| Trinity Large Preview | 83.9 |
| Sonnet 4.5 | 83.1 |
| Gemini 3 Pro Preview | 83.1 |
| Sonnet 5 | 83.1 |
| GPT-4o | 83.1 |
| Haiku 4.5 | 82.0 |
| MiMo-V2-Omni | 81.4 |
| Grok 4.3 | 80.9 |
| GLM-5.2 | 80.3 |
| MiniMax M3 | 79.2 |
| GPT-5.5 Instant | 79.0 |
| GLM-5.1 | 77.6 |
| DeepSeek-V4-Pro | 76.5 |
| GPT-4.1 | 76.5 |
| MiMo-V2.5 | 76.0 |
| DeepSeek-R1-0528 | 74.9 |
| GLM-4.7 | 73.8 |
| Qwen3.7-Max | 72.7 |
| GLM 5 Turbo | 70.5 |
| Qwen3.6 Plus Preview | 69.4 |
| Kimi K2.5 | 68.8 |
| Kimi K2.6 | 67.2 |
| MiniMax M2.1 | 66.7 |
| Qwen Max (source-unspecified mainline) | 66.7 |
| Grok 4.1 Fast | 66.1 |
| GPT-4o Mini | 65.0 |
| GPT-5.6 Sol | 63.7 |