Story
September 4, 2026
Google’s WeatherNext 3 Wants to Make the Umbrella Forecast Harder to Miss
Google presents WeatherNext 3 as a practical leap in local forecasting, especially where ground observations are sparse, while its own account acknowledges that AI is not replacing the physics-based systems agencies still rely on for warnings.
Google’s latest push into AI forecasting began with a familiar problem: traditional weather prediction is powerful but slow. Physics-based models simulate the atmosphere through complex equations, while newer AI systems search historical patterns to generate forecasts more quickly.1
With WeatherNext 3, Google says it has moved beyond that earlier AI approach by feeding the model real-time observations, including live satellite data. Samier Merchant, a Google Research engineer, said the key advance is its ability “to go beyond what data most global AI models train on” by using “fresher and richer observational data sets.”1
That shift is central to Google’s sales pitch. WeatherNext 2 issued forecasts every six hours on a 25-kilometre grid; WeatherNext 3 can generate hourly forecasts and map some variables, such as temperature and moisture, at up to five-kilometre resolution. Google says this should make fast-moving rain and snow systems easier to track, and could improve precipitation forecasts more than a day ahead by as much as 50 percent.1
The company argues the biggest gains may come in regions with fewer rain gauges — often outside the US and Europe — where satellite observations can fill holes in conventional data coverage. It is also aiming at energy markets, adding forecasts such as wind speeds at turbine height as data-centre power demand rises.1
Google DeepMind cast the release as “a major breakthrough” that learns from real-world, real-time observations to deliver more localised predictions faster.
2 WeatherNext 3 is now being integrated into Search, Maps and Gemini, and Google says it has worked with the US National Hurricane Center and agencies in Asia.
Still, the launch comes with an important caveat: WeatherNext 3 is trained partly on physics-based outputs, and weather agencies continue to compare multiple models before issuing public warnings. In other words, Google is pitching a sharper tool — not a replacement for the forecasting systems already carrying the highest-stakes calls.1