0.64
#1 of 5
safe success rate of GPT-4o, the best of five models; the best open-source model, Qwen-72B, scored 0.35
What the chart shows
ResponsibleRobotBench sets 23 multistage robot tasks involving electrical, fire or chemical, and human-related hazards. A task counts only if it is completed and every safety condition is met, so the model must spot risks, plan safe actions and ask a person for help when needed. GPT-4o scored 0.64, ahead of GPT-4o mini (0.40).
The open-source models scored lower: Qwen-72B 0.35, Qwen-7B 0.21 and InternVL 2.5 4B 0.12. Even GPT-4o failed to complete more than a third of the tasks safely.
Key numbers
| Measure | Value |
|---|---|
| GPT-4o | 0.64 |
| GPT-4o mini | 0.40 |
| Qwen-72B | 0.35 |
| Qwen-7B | 0.21 |
| InternVL 2.5 4B | 0.12 |
How to read it
- Names follow the report text; the source chart labels the models GPT4o, GPT4o-MINI, QWEN72B, QWEN7B and INTERNVL 2.5 4B.
- The CSV’s year column is each model’s release year (2023 for the Qwen models, 2024 for the others).
- Results come from one study and cover five models.
Source: Zhang et al., 2025, via the AI Index Report 2026 (Stanford HAI), figure 2.7.3: ResponsibleRobotBench: safety success rate (SSR). Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Discussion