HAL Reliability - Predictability: leaderboard
Predictability index from the Princeton HAL Reliability Dashboard (GAIA and cleaned tau-bench airline). On this dashboard it equals the inverted Brier score of the agent's own confidence against task success; calibration and AUROC are reported as sub-metrics. 0-1, higher is better.
Metric: Predictability index (0-1, higher is better). Source: hal.cs.princeton.edu. Saturation forecast: Around February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 0.84 | #87 |
| 2 | GPT-5.5 | 0.82 | #33 |
| 3 | Gemini 3.5 Flash | 0.81 | #66 |
| 4 | Gemini 3.1 Pro (Preview) | 0.8 | #64 |
| 5 | Claude Opus 4.7 | 0.8 | #51 |
| 6 | Claude Sonnet 4 | 0.79 | #207 |
| 7 | Gemini 2.5 Pro | 0.74 | #159 |
| 8 | GPT-5.2 | 0.74 | #115 |
| 9 | O1 | 0.71 | #190 |
| 10 | GPT-5.2 (Medium) | 0.67 | #115 (GPT-5.2) |
| 11 | Gemini 2.5 Flash | 0.62 | #255 |
| 12 | GPT-4 Turbo | 0.61 | #417 |
| 13 | Claude 3.5 Haiku | 0.59 | #571 |
| 14 | GPT-4o Mini | 0.46 | #598 |
| 15 | Claude 3 Haiku | 0.43 | #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-predictability · How It Works · Data refreshed daily, snapshot 2026-11-05.