Google DeepMind and Google Research introduced WeatherNext 3 on September 3, 2026, calling it their most accurate global weather model to date. The system learns from live satellite observations instead of waiting on six-hour numerical weather prediction cycles. Google said independent live evaluations by Brightband back the accuracy claim. The model is already documented for developers and shown on DeepMind’s WeatherNext page.

WeatherNext 2 produced forecasts on a 25-kilometer grid in six-hour steps. WeatherNext 3 can visualize key surface variables such as temperature and moisture at 5 kilometers, other surface variables at 10 kilometers, and atmospheric variables such as wind at 25 kilometers. Google said that is roughly five times sharper than the prior model. Forecasts now refresh every hour.

From reanalysis lag to live satellites

Most AI weather models, including WeatherNext 2, train on output from physics-based NWP systems. Those simulations are useful and slow. Google said the six-hour lag shows up as bias on fast-changing variables such as rain and surface temperature. WeatherNext 3 ingests a mosaic of live geostationary satellite data and trains directly on sparse weather-station observations, so local topography is not averaged away.

The team said that matters most in Latin America, Africa, and Asia-Pacific, where high-resolution regional models have been too expensive to run at national scale. The architecture is still a Functional Generative Network mesh transformer. It outputs dense gridded fields, discrete cyclone tracks, and station-level coordinates.

Precipitation is the usual failure mode for global models. Google trained WeatherNext 3 on NASA’s IMERG satellite retrievals and on its own satellite-radar reanalysis. Against those baselines, it reports Continuous Ranked Probability Score improvements of up to 60 percent versus IMERG, 30 percent versus MRMS, and 10 percent versus rain gauges at early lead times. The model also forecasts 100-meter wind speeds for turbine-height energy planning, plus cloud cover and surface solar radiation for solar farms.

The new forecasts start powering weather in Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine today. Google said day-ahead and longer precipitation forecasts can be up to 50 percent more accurate, with the largest gains where older products were weakest. Researchers and businesses can query the data in BigQuery and Earth Engine or bulk-download it from Google Cloud Storage. The company still tells readers to treat national weather services as the source for warnings and public safety.

Decoded Take

WeatherNext 3 is Google folding a research model into the consumer graph, then selling the same grid to energy desks and Earth Engine users. Hourly satellite updates and a 5-kilometer surface field are a real step past the 25-kilometer, six-hour cadence. The commercial question is whether a search company becomes the default weather data vendor in places that never bought a regional NWP run. Watch whether meteorological agencies cite Brightband-class scores in operations, whether the Cloud dump is current enough for grid operators, and whether Google’s own disclaimer keeps the product in the “planning aid” bucket when a storm actually hits.