Safety on GAIA
How often does the agent violate safety constraints (harmful content, policy non-compliance)? See full definition →
Sub-metric Comparison
Per-Constraint Compliance
Agent Leaderboard — Safety
| # | Agent | Acc | Safety Agg | Harm | Comp | Overall |
|---|---|---|---|---|---|---|
| 1 |  Claude 3.5 Haiku |
28.1% | 1.00 | 1.00 | 1.00 | 0.74 |
| 2 | Claude Opus 4.5 |
71.5% | 1.00 | 1.00 | 1.00 | 0.82 |
| 3 | Claude Sonnet 4.5 |
74.7% | 1.00 | 1.00 | 1.00 | 0.80 |
| 4 |  O1 |
34.7% | 1.00 | 0.75 | 1.00 | 0.72 |
| 5 |  Gemini 2.5 Flash |
37.8% | 1.00 | 0.50 | 1.00 | 0.76 |
| 6 | GPT-4 Turbo |
20.0% | 1.00 | 0.50 | 1.00 | 0.76 |
| 7 | GPT-4o Mini |
22.0% | 1.00 | 0.50 | 1.00 | 0.73 |
| 8 | Claude 3.7 Sonnet |
62.4% | 1.00 | 0.50 | 1.00 | 0.77 |
| 9 | Gemini 2.5 Pro |
50.1% | 1.00 | 0.33 | 0.99 | 0.78 |
| 10 | Gemini 2.0 Flash |
27.9% | 0.99 | 0.50 | 0.99 | 0.70 |
| 11 | GPT-5.2 |
29.9% | 0.99 | 0.53 | 0.98 | 0.74 |
| 12 | GPT-5.2 (medium) |
42.6% | 0.98 | 0.50 | 0.95 | 0.74 |