SAB Ranking
Forecasting
The ranking part of Mento.vc's Startup Agent Benchmark: 101 cohorts of 100 masked privately held companies, each ranked without tools, from descriptions and a signal panel observed at a late-2025 baseline, by likely 'operating momentum' over the next 90 and 180 days, a term the prompt leaves undefined. A ranking scores its Spearman correlation with a target that weights each company's within-cohort percentile of web-traffic growth by 0.625 and of headcount growth by 0.375, averaged over both horizons and then over cohorts; a random order scores 0 in expectation. It is a diagnostic because the target is never stated to the model and every model's correlation turns positive when the paper scores headcount growth alone.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Terra | -0.0 |
| GPT-6 Astra | -0.0 |
| GPT-6 Sol | -0.1 |
| Claude Opus 5.5 | -0.1 |
| GPT-6 Luna | -0.1 |
| Claude Fable 5.1 | -0.1 |
| DeepSeek-V3.2 | -0.1 |
| GPT-5.6 Sol | -0.1 |
| Haiku 4.5 | -0.1 |
| Claude Opus 5 | -0.1 |
| GPT-5.6 Luna | -0.1 |
| Sonnet 5 | -0.2 |