HAL Reliability - Safety: leaderboard
Safety index from the Princeton HAL Reliability Dashboard (GAIA and cleaned tau-bench airline), combining constraint compliance with the severity of harmful actions such as incorrect charges or refunds. It is not part of the overall reliability score; 0-1, higher is better.
Metric: Safety index (0-1, higher is better). Source: hal.cs.princeton.edu. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash | 1 | #66 |
| 2 | Claude Opus 4.5 | 0.99 | #87 |
| 3 | Gemini 3.1 Pro (Preview) | 0.98 | #64 |
| 4 | GPT-5.2 | 0.98 | #115 |
| 5 | GPT-5.5 | 0.98 | #33 |
| 6 | Claude Sonnet 4 | 0.98 | #207 |
| 7 | Claude Opus 4.7 | 0.97 | #51 |
| 8 | GPT-5.2 (Medium) | 0.96 | #115 (GPT-5.2) |
| 9 | Gemini 2.5 Pro | 0.96 | #159 |
| 10 | O1 | 0.95 | #190 |
| 11 | Gemini 2.5 Flash | 0.95 | #255 |
| 12 | GPT-4 Turbo | 0.94 | #417 |
| 13 | GPT-4o Mini | 0.92 | #598 |
| 14 | Claude 3.5 Haiku | 0.9 | #571 |
| 15 | Claude 3 Haiku | 0.77 | #801 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=hal-reliability-safety · How It Works · Data refreshed daily, snapshot 2026-11-05.