Google DeepMind Launches WeatherNext 3, Cutting Forecast Grid to 5km Using Live Satellite Data

What's New

Google DeepMind and Google Research released WeatherNext 3 on September 3, 2026, a global weather forecasting model the company describes as its most advanced and accurate to date. The model generates hourly forecasts at up to 5-kilometer resolution, which Google says is roughly five times sharper than WeatherNext 2, its previous flagship model that produced forecasts on a 25-kilometer grid in 6-hour increments.

The model is now integrated across Google Search, Gemini, Maps, Google Maps Platform, and Google Cloud.

Technical Details

WeatherNext 3's core change is its training data. Most AI weather models, including WeatherNext 2, learn from outputs of numerical weather prediction (NWP) systems — physics-based supercomputer simulations that carry roughly a six-hour data lag. WeatherNext 3 instead ingests a live mosaic of geostationary satellite data alongside historical analysis, feeding both into what Google calls a Functional Generative Network (FGN) mesh transformer. This architecture outputs dense gridded fields, discrete cyclone tracks, and station-level sparse coordinates directly.

Resolution varies by variable type: surface temperature and moisture are resolved at 5 kilometers, other surface variables at 10 kilometers, and atmospheric variables like wind speed at 25 kilometers. The model also trains directly on sparse weather station observations rather than only gridded atmospheric representations, which Google says improves accuracy in regions with sharp local variation — coastlines, valleys, and mountain ranges.

For precipitation — historically a weak point for both physics-based and AI weather models — WeatherNext 3 trains on NASA's satellite-based IMERG (Integrated Multi-satellite Retrievals for GPM) dataset combined with Google's own precipitation data.

The model adds forecasting for clean-energy applications: 100-meter wind speed predictions for turbine-height wind-energy output, and high-resolution cloud cover and solar radiation estimates for solar farm output planning.

Google cites independent live evaluations by Brightband as the basis for its claim that WeatherNext 3 is the most accurate global weather model available — this is a third-party claim, not an internally verified benchmark disclosed with specific scores.

Pricing and Availability

Pricing for direct API or Cloud access has not been disclosed. The model is available now through consumer-facing Google products (Search, Maps, Gemini) and via Google Cloud and Google Maps Platform for developers.

What This Means

WeatherNext 3 shifts the underlying data paradigm for AI weather forecasting — away from NWP-simulation outputs and toward live satellite observations — which directly addresses the 6-hour lag problem that has limited AI models' ability to track fast-developing storms and precipitation. The jump from 6-hour to hourly forecast cycles, combined with 5-kilometer resolution, is a meaningful technical improvement for use cases like storm response, agriculture, and renewable energy grid planning, particularly in regions across Latin America, Africa, and Asia-Pacific that previously lacked high-resolution regional forecasting due to supercomputing costs.

The accuracy claims rest on third-party evaluation by Brightband rather than a standardized, widely-adopted benchmark suite, so independent replication by other forecasting groups will matter for validating the five-times-sharper claim. Google's decision to route WeatherNext 3 immediately into Search, Maps, and Gemini means the model's real-world performance will be tested at consumer scale almost immediately, rather than through a slower research-to-product pipeline.