Benchmarks / Agentic
HiL-Bench
Uses Scale Labs combined Pass@3 with ask_human tool. Score: combined Pass@3 with ask_human tool. Collected by Jake Boggs' benchmark dashboard.
HiL-Bench
Higher is better · best effort per model
full table below
| # | Model | Effort | Score |
|---|---|---|---|
| 1 | Fable 5.1 | max | 61.5% |
| 2 | Opus 5 | max | 57.0% |
| 3 | Fable 5 | max | 56.3% |
| 4 | GLM 5.2 | max | 43.7% |
| 5 | Opus 4.7 | max | 41.7% |
| 6 | Gemini 3.8 Flash | max | 41.5% |
| 7 | GPT-5.5 | xhigh | 39.7% |
| 8 | Opus 4.6 | max | 38.3% |
| 9 | Opus 4.8 | max | 35.3% |
| 10 | Gemini 3.1 Pro | high | 35.3% |
| 11 | GPT-5.6 Sol | max | 32.3% |
| 12 | Gemini 3.5 Flash | xhigh | 27.7% |
| 13 | Grok 4.20 | xhigh | 20.0% |
| 14 | Kimi K2.6 | max | 18.7% |
| 15 | GPT-5.4 | xhigh | 9.7% |
| 16 | MiniMax M2.5 | high | 6.3% |
| 17 | GPT-5.3 Codex | xhigh | 4.3% |