Google’s new AI predicts rain 50% better — it’s already in Maps and Search

The weather forecast on your phone just got sharply more reliable without you doing anything. Google DeepMind replaced the model behind its weather cards in Search, Maps, Gemini, and the Maps Platform API with WeatherNext 3 — a system that produces predictions every hour at 5-kilometer resolution, compared to six-hour intervals on a 25-kilometer grid from its predecessor.

The practical difference shows up most clearly in rain. For planning a day or more ahead, WeatherNext 3’s precipitation forecasts are up to 50 percent more accurate than the previous model’s. Against IMERG, the satellite benchmark used by climate researchers, the improvement reaches 60 percent. That translates to the difference between a forecast that says ‘possible showers’ and one that tells you to expect rain between 2 and 5pm in your specific neighborhood.

WeatherNext 3 also outperforms traditional numerical weather prediction systems operated by the U.S. National Weather Service and European meteorological agencies, according to Google’s own benchmarks. The model extends beyond rain: it includes wind speeds at 100 meters — the height of a wind turbine — cloud cover, and solar radiation levels, making it useful for energy operators and agricultural planners alongside ordinary commuters.

Google is not the only company developing AI-based weather models. Microsoft, Nvidia, and European research institutions have published competing approaches in the past two years. Google’s announcement did not name specific rivals, and independent meteorological agencies have not yet published external validation of WeatherNext 3’s claimed accuracy gains — the figures come from Google’s own comparative testing.

The key architectural change is how the model receives data. Traditional weather forecasting systems collect observations over a six-hour window before running their calculations. WeatherNext 3 ingests live geostationary satellite imagery every hour, bypassing that lag. It also trains on direct readings from ground-level weather station networks, which helps it reflect local terrain — coastal effects, mountain passes, and urban heat signatures that a coarser grid would smooth away.

There are real limits to what 5-kilometer resolution means in practice. The model cannot resolve the difference between rain hitting one side of your street before the other, and it cannot override the fundamental unpredictability of convective weather systems. Google’s own blog directs users to local meteorological agencies for severe weather warnings and public safety advisories.

Pre-trained versions of WeatherNext 3 are available to researchers through Google Cloud and Google Earth Engine as of September 3, 2026, when the model was announced. No phased regional rollout was announced — the upgrade is live globally for all users of Google’s consumer weather products.

Tags: Google, ai, Machine Learning, DeepMind, Google Maps, weather forecast