HAL Reliability - Robustness: leaderboard
Robustness index from the Princeton HAL Reliability Dashboard (GAIA and cleaned tau-bench airline): the mean of fault-injection tolerance, structural-perturbation robustness and prompt-rewording robustness. Fault and structural robustness sit near the ceiling for most agents; 0-1, higher is better.
Metric: Robustness 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 | O1 | 0.99 | #190 |
| 2 | Gemini 3.5 Flash | 0.97 | #66 |
| 3 | GPT-5.5 | 0.96 | #33 |
| 4 | Claude Sonnet 4 | 0.95 | #207 |
| 5 | GPT-4o Mini | 0.95 | #598 |
| 6 | Claude Opus 4.7 | 0.95 | #51 |
| 7 | Claude Opus 4.5 | 0.94 | #87 |
| 8 | Gemini 3.1 Pro (Preview) | 0.94 | #64 |
| 9 | Claude 3.5 Haiku | 0.93 | #571 |
| 10 | GPT-5.2 (Medium) | 0.93 | #115 (GPT-5.2) |
| 11 | Claude 3 Haiku | 0.92 | #801 |
| 12 | Gemini 2.5 Pro | 0.92 | #159 |
| 13 | GPT-5.2 | 0.85 | #115 |
| 14 | Gemini 2.5 Flash | 0.85 | #255 |
| 15 | GPT-4 Turbo | 0.84 | #417 |
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-robustness · How It Works · Data refreshed daily, snapshot 2026-11-05.