Proof of Time: leaderboard
Forecasting judgments about research that can only be checked later (future citations, award tiers, researchers' next topics, benchmark progress), answered by a tool-using ReAct agent working offline over a frozen pre-cutoff evidence snapshot with a 50-message limit. Sample-weighted exact-match accuracy (%), higher is better (Proof of Time, Ye, Chen et al., Harvard University and Mass General Brigham, 2026).
Metric: Sample-Weighted Accuracy (%), 50-message agent budget. Source: arxiv.org. Saturation forecast: Around October 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 55.8 | #111 |
| 2 | Gemini 2.5 Pro | 53.6 | #262 |
| 3 | GPT-5.1 | 45 | #225 |
| 4 | Claude Opus 4.5 | 41.2 | #152 |
| 5 | GPT-5.2 | 40.1 | #172 |
| 6 | Gemini 3 Flash (Preview) | 39.2 | #141 |
| 7 | GPT-5 Mini | 37.7 | #349 |
| 8 | Gemini 2.5 Flash | 37.5 | #452 |
| 9 | Claude Haiku 4.5 | 35.1 | #568 |
| 10 | GPT-5 Nano | 33.9 | #888 |
| 11 | Claude Sonnet 4.5 | 31 | #244 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=proof-of-time · How It Works · Data refreshed daily, snapshot 2026-11-12.