Google WeatherNext 3 Adds Satellite Data, Boosts Forecast Accuracy
Google’s WeatherNext 3 adds satellite data and hourly forecasts, improving upper-air accuracy 5% and local surface temperatures by up to 30% across Google.
Summary
Google released WeatherNext 3 on September 8, 2026, adding satellite weather observations to the reanalysis data used by most AI forecasting systems. Reanalyses merge global measurements and estimates into snapshots generally produced every six hours, potentially losing detail from raw sources. Satellite ingestion reduces that delay and enables hourly forecasts. WeatherNext 3 also increases spatial resolution and model size, with process changes limiting added computing demand, while a separate model trained on satellite precipitation estimates produces additional rainfall forecasts. For location-specific surface temperature and dew point, it incorporates land or ocean status, elevation and tagged weather-station history. Like other pattern-based AI forecasters, it requires far less computing than traditional physics models and can run more frequently.
Google reports roughly 5 percent better upper-atmosphere accuracy than WeatherNext 2, equivalent to about six additional hours of accurate lead time, and up to 30 percent better location-specific surface-temperature accuracy. WeatherNext 3 generally beats the European Centre for Medium-Range Weather Forecasts AI model on those measures, but several variables perform worse at the initial six-hour comparison before pulling ahead during the rest of the 15-day forecast. Remaining anomalies include hexagonal grid patterns in precipitation maps and forecast sets whose global average temperatures drift higher or lower instead of preserving a stable mean. WeatherNext 3 now supplies forecasts across Google Search, Gemini and Maps.
Positives
- Hourly forecasts replace the six-hour cadence enabled by reanalysis-only inputs after WeatherNext 3 adds lower-latency satellite observations.
- Upper-atmosphere accuracy improves roughly 5 percent over WeatherNext 2, providing about six additional hours of accurate forecast lead time.
- Location-specific surface-temperature accuracy rises by up to 30 percent after adding land or ocean status, elevation and weather-station training.
- WeatherNext 3 generally outperforms the European Centre for Medium-Range Weather Forecasts AI model on the measured accuracy metrics.
- Google Search, Gemini and Maps now use WeatherNext 3 forecasts, extending the improvements across widely used consumer services.
Risks & concerns
- Several forecast variables underperform rival models at the initial six-hour horizon before WeatherNext 3 pulls ahead later in the 15-day period.
- Hexagonal grid artifacts remain visible in some precipitation predictions, exposing structural irregularities in the model’s output.
- Multiple surface-temperature forecasts sometimes shift the global average higher or lower instead of producing balanced local variation.
- The larger model and higher spatial resolution increase computational demands, although Google introduced process changes to limit the impact.