HiL-Bench: leaderboard
Scale Human-in-Loop benchmark measuring when agents should ask for help, escalate uncertainty, or continue autonomously.
Metric: Combined Pass@3 (%). Source: labs.scale.com. Status: years away from saturation. 15 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 5 | 57 |
| 2 | Claude Fable 5 | 56.33 |
| 3 | GLM-5.2 | 43.67 |
| 4 | Claude Opus 4.7 | 41.67 |
| 5 | GPT-5.5 | 39.67 |
| 6 | Claude Opus 4.6 | 38.33 |
| 7 | Gemini 3.1 Pro (Preview) | 35.33 |
| 8 | Claude Opus 4.8 | 35.33 |
| 9 | GPT-5.6 Sol | 32.33 |
| 10 | Gemini 3.5 Flash | 27.67 |
| 11 | Grok 4.20 | 20 |
| 12 | Kimi K2.6 | 18.67 |
| 13 | GPT-5.4 | 9.67 |
| 14 | MiniMax-M2.5 | 6.33 |
| 15 | GPT-5.3 Codex | 4.33 |
Interactive version: theaggregate.ai/benchmark?slug=hil-bench · How It Works · Data refreshed daily, snapshot 2026-09-05.