WeatherNext 3 Brings Hourly, 5 km AI Weather Forecasts

Does a weather model that updates every hour instead of every six change anything for a business? For most offices, no. For a wind farm bidding tomorrow's output into a power market, a cold-chain carrier routing trucks over a mountain pass, or a grower deciding which morning is calm enough to spray, the answer Google DeepMind is betting on with WeatherNext 3, announced September 3, is yes. It is the first global model with hourly updates, and it pushes surface temperature and humidity forecasts to a 5-kilometer grid.

Three changes from WeatherNext 2

The previous model, released last year, refreshed every six hours on a 25-kilometer grid. The new one changes the cadence to hourly, sharpens resolution to 5 km for temperature and humidity and 10 km for other surface variables such as wind, and, most interestingly, changes what it eats. Rather than waiting for the processed output of conventional numerical weather prediction systems, it ingests raw imagery from geostationary satellites directly, as hourly mosaics. That removes a lag of roughly six hours built into the old pipeline. Google describes the result as "roughly five times sharper." New variables target renewable energy specifically: wind speed at 100 meters (turbine hub height), high-resolution cloud cover, and solar radiation. It remains an ensemble model, producing a spread of scenarios rather than a single line.

Reading the benchmark claims carefully

Google's precipitation numbers: improvements of up to 60 percent against NASA's IMERG satellite dataset, 30 percent against the radar-based MRMS system, and 10 percent against rain gauges at short lead times, with "up to 50 percent" gains at longer ranges and the biggest wins in regions where forecasting infrastructure is weakest. Each figure is relative to a specific reference dataset, and the announcement does not include a head-to-head comparison with the European Centre for Medium-Range Weather Forecasts, the benchmark Google used for earlier versions. For independent verification the company points to live leaderboards run by Brightband. It also adds the standard caveats: the atmosphere "will always retain a degree of unpredictability," and official warnings should come from national weather services.

Where businesses can get it

Consumers will see it in Google Search, the Gemini app, and Google Maps. The business routes are the Google Maps Platform Weather API, Google Earth Engine, a public dataset in BigQuery, and bulk download from Google Cloud Storage, with documentation at developers.google.com/weathernext. No WeatherNext 3-specific pricing was published; it appears to sit under existing Cloud and Maps Platform pricing, though the announcement does not say so explicitly.

Our take

The interesting shift is not the accuracy percentage but the unit of analysis: from a city's daily forecast to a specific parcel at a specific hour. That is the granularity operational decisions actually run on, and until now it was available mainly to organizations that could afford custom meteorology. The BigQuery route means a planning team can join the forecast to its own schedule table without training anything. Our advice is to spend the first quarter comparing the model's output against your own local measurements before wiring it into decisions. A global 60 percent improvement is not a promise about your valley.

Sources: Google announcement, Google DeepMind WeatherNext page