ForecastBench: leaderboard
Measures LLM ability to forecast future events - think prediction markets. 1,000 auto-generated questions from Polymarket, Manifold, and real-world data. Compares against superforecasters and the public.
Metric: Overall Score (higher is better). Source: www.forecastbench.org. Status: years away from saturation. 299 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Median Expert | 70.7 |
| 2 | Grok 4.20 | 67.9 |
| 3 | GPT-5 | 65.7 |
| 4 | Gemini 3 Pro (Preview) | 65.6 |
| 5 | Grok 4 | 65.3 |
| 6 | Grok 4 Fast (Reasoning) | 65.3 |
| 7 | Claude 3.7 Sonnet (20250219) | 65.2 |
| 8 | GPT-4.5 (Preview) | 65.2 |
| 9 | Grok 4.1 Fast (Reasoning) | 65.1 |
| 10 | Claude Opus 4.1 (20250805) | 65.1 |
| 11 | Claude Opus 4.5 (20251101) | 65.1 |
| 12 | Median Human | 65.1 |
| 13 | Gemini 3.1 Pro (Preview) | 64.9 |
| 14 | Gemini 3 Flash (Preview) | 64.9 |
| 15 | O3 (2025-04-16) | 64.9 |
Interactive version: theaggregate.ai/benchmark?slug=forecastbench · How It Works · Data refreshed daily, snapshot 2026-09-05.