On August 6, 2026, Google DeepMind published research in Nature reporting state-of-the-art accuracy in predicting a tropical cyclone’s track, intensity, and wind structure. The headline result is lead time: DeepMind says its three-day forecasts are as good as what prior models could provide for only the next two days, an average gain of more than a full 24 hours. The company characterizes that jump as roughly a decade’s worth of meteorological progress measured against the trend of the last 20 years.
The model uses Functional Generative Networks to produce ensembles of predictions efficiently, and it runs at a 28 by 28 kilometer resolution that DeepMind notes is about 100 times coarser than traditional physics-based models while still outperforming them. Ensembles ran at 50 members at a time during the 2025 hurricane season and scale to 1,000 members for capturing rare but consequential scenarios. DeepMind released code and model weights on GitHub for academic research, operational forecasting, and building more specialized or localized models.
The model has already been used operationally rather than only in retrospective evaluation. DeepMind says the US National Hurricane Center used it during Hurricane Melissa in 2025, where it contributed to predicting the storm’s rapid intensification and its landfall in Jamaica. That is the meaningful distinction from a benchmark claim: a national forecasting agency ran it against a live storm.
For decision-makers the unit that matters is the extra day, because cyclone response is a sequence of deadlines rather than a single prediction. Evacuation orders, offshore shutdowns, grid staging, and supply repositioning all have lead times measured in hours, so a day of additional warning changes what is physically possible before landfall rather than merely improving a score. The open release also means the capability is not gated behind one provider, which matters for national meteorological services that cannot outsource statutory forecasting duties.