WeatherNext 3 Adds Hourly Forecasts and Selected 5 km Fields
Google announced hourly WeatherNext 3 forecasts, with selected surface variables at 5 km resolution, beginning rollout across Search, Gemini and Maps. Google cites Brightband’s evaluation for its leading-accuracy claim.
Correction / update, October 14, 2026: Only selected surface variables use a 5 km grid. The model also uses traditional analysis; rollout and the Google-relayed Brightband ranking are now qualified.
Weather is one of the few AI products users feel daily. An hourly model with selected 5 km fields, beginning rollout across Search and Maps, extends Google’s reach.
Google DeepMind and Google Research (September 3) published WeatherNext 3. It uses live geostationary satellites and station observations alongside traditional analysis. It provides hourly forecasts, compared with WeatherNext 2’s six-hour intervals. Selected surface variables use 5 km grids, five times finer than the prior 25 km grid; other surface variables use 10 km grids and atmospheric fields remain at 25 km. Precip CRPS up to 60% vs IMERG, 30% MRMS, 10% gauges at early leads; consumer surfaces “up to 50% more accurate” precip, biggest gains in historically weak regions. Adds 100 m winds + solar radiation for renewables. Google cites Brightband live evaluations for its most-accurate-global-model claim; the ranking was not independently verified here. Rollout began across Search, Gemini, Maps, Maps Platform Weather API and Earth Engine, with forecast data accessible through BigQuery, Earth Engine and GCS. Paper: arXiv:2609.03582.
WeatherNext 3 is a science and weather model, rather than a conversational model.
Limits#
- The live Brightband leaderboard was not independently inspected for this article; Google’s ranking claim is attributed, not a verified current rank.
- Consumer “up to 50% more accurate” is Google’s framing of precip gains.