2026 hurricane season: DeepMind's model outperforms the U.S. National Center by up to 30% at five days

The verifications compiled by a former official of the National Hurricane Center place Google's DeepMind system ahead of all other public track forecasts, in the Atlantic as well as in the Pacific.

For the second year in a row, Google DeepMind's hurricane forecasting model dominates track comparisons. The verification data collected by James Franklin, former head of the National Hurricane Center's tropical storm forecasting branch, and as of September 6, 2026, puts it ahead of all the public aid available in the Atlantic, East Pacific and Central Pacific basins.

The gap widens especially at longer time frames, where forecasting is the most difficult: up to 30% less error than the American center on a five-day forecast. In terms of intensity, on the other hand, the system is on par with the NHC and remains behind HCCA, an adaptive consensus aid that retains the advantage in this category.

The European AI model, which is promising in 2025, suffered from significant errors in intensity until a recent correction. The hierarchy that is emerging is therefore nuanced: machine learning has taken a clear lead on the trajectory, without having yet resolved the issue of intensification.