Bench to the Future 2
Forecasting · 2026-04-20
Bench to the Future 2 (BTF-2) evaluates forecasting agents on 1,417 resolved pastcasting questions researched against a frozen 15-million-document corpus. Its headline metric is mean Brier score, so lower values indicate more accurate probabilistic forecasts.
Top models (lower is better)
| Model | Score |
|---|---|
| Opus 4.6 | 0.1 |
| Gemini 3.1 Pro Preview | 0.1 |
| GPT-5.4 | 0.2 |
| Grok 4.20 | 0.2 |