Google WeatherNext 3 Beats Leading AI and Government Forecasts
Google WeatherNext 3 brings 5 km hourly forecasts and 60% better rain scores to Search, Maps and Gemini after beating leading weather models in benchmark tests.
Summary
On September 3, 2026, Google DeepMind and Google Research released WeatherNext 3, which will supply weather variables to Google Search, Maps and Gemini and be available through Google Cloud. Senior staff engineer Samier Merchant said this is the first time core variables will power many Google products. The Brightband startup’s Operational WeatherBench ranked it the most accurate leading model tested for temperature, wind speed and humidity, ahead of AI systems from Google, Microsoft, Nvidia and the European Center for Medium-Range Weather Forecasting, plus conventional US National Weather Service and ECMWF forecasts.
WeatherNext 3 resolves key variables at 5 kilometers, versus the 15 to 25 square kilometer areas typical of AI models, improves rain evaluations 60% over WeatherNext 2 and forecasts hourly rather than every six hours. It has 2.4 times more parameters, tailored decoder targets, cyclone path visualization and forecasts tied to individual stations. Brightband atmospheric scientist Daniel Rothenberg said predicting hourly readings at sites such as Denver airport connects forecasts more closely to ground truth. Hourly satellite ingestion enables more frequent forecasts from raw observations.
Google calls WeatherNext 3 the first AI model to directly incorporate raw observations into high resolution global forecasts. WindBorne says WeatherMesh 6 has used observations from balloons and other sources since late 2025, while Google says its global resolution is higher. Both still require national weather datasets, leaving true direct data assimilation unfinished. ECMWF’s 2018 release of more than 50 years of data helped AI models challenge slower, costlier supercomputer forecasts. European and US agencies now use AI forecasting. Bill Gates has linked better forecasts to developing world crop yields, while DeepMind’s Ferran Alet says finer wind, rain and cloud predictions could improve renewable energy reliability and extend affordable forecasting to poorer regions.
Positives
- WeatherNext 3 ranked first on Brightband’s Operational WeatherBench for temperature, wind speed and humidity.
- Five kilometer resolution gives WeatherNext 3 substantially finer predictions than the 15 to 25 square kilometer areas typical of AI models.
- Rain evaluations improved 60% over WeatherNext 2, while forecast frequency increased from every six hours to hourly.
- Google plans to integrate WeatherNext 3 variables into Search, Maps and Gemini and offer access through Google Cloud.
- Hourly satellite ingestion and station level targets connect forecasts more directly to current observations and measurable ground truth.
- Faster, cheaper AI forecasts could support developing world crop yields, renewable energy reliability and regions lacking costly supercomputers.
Risks & concerns
- WeatherNext 3 and WindBorne’s WeatherMesh 6 still depend on national weather datasets, leaving true direct data assimilation unfinished.
- WindBorne contests Google’s first model claim, citing raw observations from balloons and other sources in WeatherMesh 6 since late 2025.
- Processing unformatted raw observations remains technically challenging despite WeatherNext 3’s direct satellite data ingestion.
- High quality sensors and supercomputers remain too expensive for many poorer regions, restricting access to conventional accurate forecasting.