How Oracle predicts.
Oracle helps you turn a question about the future into something specific. Then the Oracle researches the evidence and estimates how likely that outcome is.
Start with your question.
You don’t need perfect wording. Text what’s on your mind. Oracle asks follow-ups to understand your situation, then drafts a question with a clear outcome and deadline. Tweak it until you’re happy. Research starts after you confirm.
Follow the evidence.
The Oracle researches the question and weighs the case for and against. You get a probability, the reasoning behind it, and sources you can follow.
The published track record.
The homepage chart shows published ForecastBench scores, checked October 13, 2026. The forecasting score (Overall Brier Index) runs from 0 to 100, with higher scores better. It measures forecast quality, not the percentage of questions answered correctly. The chart’s vertical axis runs from 55 to 70.
The human and standalone-model bars come from ForecastBench’s baseline leaderboard. These models are tested without added tools, context or research scaffolding. “Avg human” is the public median forecast reference, surveyed in July 2024, scoring 62.5. “Claude” is Claude Sonnet 5 with adaptive thinking and a 16,000-token budget (claude-sonnet-5-adaptive-thinking-16000), scoring 60.4. “ChatGPT” labels the GPT-5.5 model (gpt-5.5-2026-04-23), scoring 60.8. These are the tested model versions, not direct measurements of today’s default ChatGPT or Claude apps. Inspect the baseline scores.
The starred “the Oracle” bar shows FutureSearch’s futuresearch_bull_v4 tournament result of 68.5 as an engine reference. The tournament allows research tools and other enhancements. This is a separate evaluation with a different question mix and difficulty adjustment, so the four bars are not a head-to-head comparison. The human reference in that tournament scores 63.4. Inspect the tournament scores or read the methodology.
The baseline scores have overlapping 95% confidence intervals: human reference 60.0–65.3, Claude 59.2–61.6 and GPT-5.5 59.9–61.7. Their point estimates do not establish that today’s chat apps are generally worse than humans. FutureSearch’s tournament interval is 67.2–69.9. Oracle uses FutureSearch’s faster forecast API; its results can differ from the tournament system’s. The published tournament score does not measure Oracle’s own accuracy or guarantee an individual prediction.
Understand the odds.
A prediction is an estimate, not a promise. A 70% chance means an outcome is more likely than not, but it can still go the other way. The explanation helps you understand the assumptions and what could change the answer.
Keep the useful context.
The facts you share help Oracle understand your situation. It can research public information, but it can’t know your private circumstances unless you tell it. Share what’s relevant to the question.