RSO-F ranks above OpenAI, Anthropic and xAI on ForecastBench (September 2026)

RSO-F has made a strong start on ForecastBench, a leading independent benchmark for evaluating how well AI models predict real-world events. In the 22 September 2026 preliminary dataset results, our first submission scored above the highest-scoring listed entries from OpenAI, Anthropic, xAI and FutureSearch.

Listed as RSO-F-ST, the model scored 64.9 on the Dataset Brier Index across 241 resolved questions from the seven-day horizon of our 30 August submission. That places RysoAI close to Google DeepMind and Google Cloud at 65.0, with Torchcast AI at 65.8 in this comparison.

What this means

This is an early demonstration of what we are building RSO-F to do: turn information available today into better predictions of what happens next. A first public submission that competes with established forecasting teams and outperforms the compared frontier-model entries gives us a strong foundation to build on.

Joint first on conflict and protest data

RSO-F ranked joint first on ACLED in the original seven-day results. Across the 50 ACLED questions from our 30 August submission, RSO-F-ST scored 0.0487 mean Brier using the outcomes archived on 22 September 2026. It tied with our two other Ryso variants at the top of 112 eligible entries, ahead of the best listed configurations from OpenAI, Anthropic, xAI, Google DeepMind, Google Cloud and Torchcast AI.

ACLED records political violence and protest events around the world. These ForecastBench questions ask whether event counts or fatalities in a rolling 30-day window will exceed thresholds derived from historical data, with the outcome assessed seven days after submission. The result points to a promising capability for forecasting changes in conflict and protest activity.

Update, 10 October 2026: one outcome was subsequently revised. Recalculating the same forecasts gives RSO-F a 0.0685 Brier score and joint sixth place. The chart preserves the original September result; the downloadable data includes both versions.

Reading the results

The first chart captures ForecastBench’s preliminary dataset leaderboard on 22 September 2026. It shows the highest-scoring listed entry for each selected organization. RSO-F’s result comes from its first seven-day forecasts; the other entries have different question sets and submission histories.

The Dataset Brier Index measures probability forecast performance after adjusting for question difficulty; higher is better. Lines show 95% confidence intervals. The preliminary board covers dataset questions; the tournament leaderboard combines dataset and prediction-market results.

Where we need to improve

The short-term ACLED result is stronger than the longer-horizon result. On the separate 50-question, 30-day set from the same round, RSO-F-ST ranks 29th of 112 entries in our 10 October calculation. Part of the rolling 30-day observation window has already elapsed when a seven-day forecast is submitted. Forecasting further ahead demands more of the model, and improving that performance is a priority.

The next test

We expect some performance decay as longer-horizon outcomes are scored. We also expect RSO-F to remain highly competitive. The next milestone is the combined tournament leaderboard: ForecastBench applies a 50-day waiting period before eligibility, putting our 30 August submission’s earliest entry on 19 October. We will publish the full dataset and prediction-market comparison when it appears.

For the two submission rounds following our 30 August debut, we are testing a range of approaches and specific forecasting cases against live data. We expect these experimental submissions to perform roughly in the middle of the pack while we learn what works. Each test helps us understand where the model is strong, where it needs work and which improvements carry over to new questions.