Google DeepMind announced WeatherNext 3 on September 3, 2026, describing it as its most advanced and accurate global weather model. The headline change is resolution and cadence. Where WeatherNext 2 produced forecasts on a 25 km grid in six-hour increments, WeatherNext 3 updates hourly and resolves surface variables such as temperature and moisture down to 5 km, other surface variables to 10 km, and atmospheric variables such as wind speed to 25 km. Google describes the global picture as roughly five times sharper than the previous version while maintaining physical consistency from broad global wind patterns down to local topography.
The accuracy claims are stated against conventional observational baselines rather than against other models alone. For precipitation, Google reports a Continuous Ranked Probability Score improvement of up to 60 percent against IMERG, 30 percent against MRMS, and 10 percent against rain gauge measurements at early lead times, and up to 50 percent more accurate precipitation forecasts for longer-range planning.
What makes this a deployment story rather than a research story is where the model runs. WeatherNext 3 is already powering weather information in Google Search, Google Maps, and the Gemini app, and is available to builders through the Google Maps Platform Weather API, Google Earth Engine, BigQuery, and Google Cloud Storage. That is a neural weather model serving consumer-scale traffic as the default, not an experimental overlay.
For leaders outside meteorology, this is the clearest live example of AI displacing a mature numerical simulation pipeline in production. National weather services spent decades building physics-based forecasting on supercomputers; a learned model now beats parts of it on skill, runs far cheaper, refreshes hourly, and ships inside a mapping app. The same displacement argument gets made about every simulation-heavy field, but here it has stopped being an argument.