TextQuests Harm per 100 Steps
Safety · 2025-07-31
TextQuests evaluates LLM agents playing 25 classic interactive-fiction games. This row reports harm per 100 steps: impermissible actions such as violence or theft normalized by trajectory length; lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-5 Nano | 2.0 |
| Grok 4.3 | 3.9 |
| Grok 4.1 Fast | 9.1 |
| Grok 4.20 | 9.9 |
| Grok 4.5 | 11.2 |
| GPT-4o (2024-11-20) | 11.2 |
| Gemini 2.5 Flash | 11.7 |
| o3-mini | 13.4 |
| Kimi K2 Thinking | 14.6 |
| GPT-5.4 Nano | 15.2 |
| Muse Spark 1.1 | 15.2 |
| DeepSeek-R1 | 15.4 |
| GPT-5.4 Mini | 15.4 |
| Gemini 3.1 Flash-Lite Preview | 15.6 |
| Gemini 2.5 Pro | 15.9 |
| Sonnet 4 | 16.0 |
| Sonnet 4.6 | 16.5 |
| Haiku 4.5 | 16.6 |
| GPT-5.1 | 16.7 |
| Kimi K3 | 17.0 |
| Sonnet 4.5 | 17.1 |
| GPT-5 | 17.2 |
| GPT-5 Mini | 17.5 |
| Gemini 3 Flash Preview | 17.8 |
| GLM-5.2 | 18.1 |
| O4 Mini | 18.6 |
| o3 | 18.7 |
| Gemini 3.1 Pro Preview | 18.8 |
| Grok 4 Fast | 19.8 |
| Opus 4.5 | 20.3 |
| Kimi K2.6 | 20.4 |
| GPT-5.2 | 20.5 |
| Gemini 3.5 Flash | 20.7 |
| Kimi K2.5 | 20.8 |
| Opus 4.8 | 21.0 |
| Gemini 3 Pro Preview | 21.4 |
| GPT-5.5 | 21.4 |
| DeepSeek-V3.2 | 21.6 |
| Opus 4.7 | 21.8 |
| GLM-5.1 | 22.0 |