AI Reaches Human SuperForecaster Range

Superforecasters — the small group of people who consistently outpredict everyone else — have held a clear lead over language models on the hardest forecasting benchmarks. This gap has been one of the more durable results in AI evaluation.

On ForecastBench, that gap is beginning to close.

ForecastBench is a continuously updated forecasting benchmark from the Forecasting Research Institute. Instead of asking models questions with known answers, it poses questions about events that haven't happened yet, then scores each forecaster once the outcomes are in. Because the questions resolve in the real world, there is nothing to memorize and nothing to look up. It measures reasoning about an uncertain future — which is exactly what forecasting is.

While there are fair questions about whether ForecastBench is perfectly comparing human superforecasters to AI (ForecastBench adjusts for question difficulty because its human panel was last surveyed in 2024 on a different question set), we believe it’s the best source to date for evaluating AI against human superforecasting capability.

On this week's tournament leaderboard, the human superforecaster median still ranks #1 (Overall Brier Index 69.3). But the models directly beneath them are no longer trailing at a distance. Several — including Google's Gemini, an ensemble system called Cassi, and Foresight-v3 — land inside the superforecasters' confidence interval. ForecastBench reports this directly, in a column that asks whether the superforecasters are significantly better than each forecaster. For the top handful of AI systems, the answer is now "no."

For the first time on ForecastBench, the best AI forecasters are statistically indistinguishable from the best humans. 

Our previous-generation model, Foresight-v3, reached the #4 rank with an overall Brier Index score of 67.7. While rankings on the board move daily, it’s worth noting that among the models at this level, Foresight-v3 is the only one that:

We’re excited to see a small model perform so strongly alongside much bigger models and purpose-built harnesses, as well as alongside human superforecasters. 

Note: Foresight-V4 has been submitted for upcoming ForecastBench competitions.

With AI Forecasters this close to human superforecaster performance, it feels like we are approaching a critical moment in AI capability. Seeing these results from both large frontier models and small, domain-specific models points to a future where AI prediction can be applied to many use cases across public and private data sets alike. 

For fun, let’s end with a prediction from Foresight V4: