WeatherNext 3: Google's weather AI now refreshes hourly
Hourly runs at 5-kilometer resolution, satellite data fed straight in, and a rollout across Search, Gemini and Maps since September 3.
Symbolic image: at sunrise a technician mounts a spinning wind sensor on an instrument mast while the status lights of a datalogger blink alongside.
WeatherNext 3 is Google's new global AI weather model, producing a fresh forecast every hour at resolutions down to 5 kilometers by reading live satellite imagery directly.
At a glance
- Tiered resolution: 5 km for surface variables such as temperature, 10 km for other surface fields, 25 km for atmospheric variables.
- Roughly five times sharper than WeatherNext 2, which ran on a 25 km grid in six-hour steps.
- Precipitation: up to 60% better CRPS against NASA IMERG, 30% against MRMS, 10% against rain gauges at short lead times.
- Live since September 3, 2026 in Search, the Gemini app, Maps and the Maps Platform Weather API; data via BigQuery and Earth Engine.
- New additions: wind speed at 100 meters altitude plus cloud cover and solar radiation for wind and solar farms.
WeatherNext 3 is Google's new global AI weather model. It produces a fresh forecast every hour, resolves surface variables down to 5 km, and takes satellite imagery in directly rather than leaning solely on conventional numerical weather prediction. Google and Google DeepMind announced it on September 3, 2026.
Five times sharper than the last one
Resolution is tiered: 5 km for surface variables such as temperature and moisture, 10 km for other surface fields, and 25 km for atmospheric variables like wind. Its predecessor, WeatherNext 2, ran on a 25 km grid in six-hour steps. Google puts the jump at roughly five times sharper.
Under the hood the team describes a Functional Generative Network built as a mesh transformer. It emits more than gridded fields — discrete cyclone tracks and values at the coordinates of individual weather stations come out of the same model.
Rainfall is the real pitch
Training drew on NASA's IMERG record and on a global precipitation reanalysis of Google's own. Measured against IMERG at medium range, Google reports a CRPS improvement of up to 60%; against the MRMS radar product it claims 30%, and against rain gauges 10% at early lead times. For anyone planning a day or more out, the post cites precipitation forecasts up to 50% more accurate.
Where the forecasts surface
Consumers meet the model in Google Search, the Gemini app, Google Maps and Earth Engine, and developers reach it through the Maps Platform Weather API. Bulk data flows via BigQuery and Google Cloud Storage. For energy operators, the model adds wind speed at 100 meters altitude along with cloud cover and solar radiation.
What this story does not establish
Every figure here comes from a single place: the vendor's own blog post. No second, independent newsroom had confirmed the numbers at the time of reporting, and we have no outside re-measurement.
Also unstated: how many days ahead the forecast reaches, how large the ensemble is, what a run costs in compute and energy, and how well tropical cyclone tracks are hit — the post gives no values for any of it. Google names Brightband as an independent evaluator but does not quantify what that evaluation found. For official warnings, Google points readers back to their national meteorological service.
FAQ
What is WeatherNext 3?
A global AI weather model from Google and Google DeepMind, announced on September 3, 2026. It issues a new forecast every hour, resolves surface variables down to 5 km, and continuously ingests satellite data.
Where can I see WeatherNext 3 forecasts?
In Google Search, the Gemini app and Google Maps. Developers can pull the data through the Maps Platform Weather API, Earth Engine and BigQuery.
Does WeatherNext 3 replace an official weather service?
No. Google states in the post that official forecasts, severe weather warnings and public safety advisories remain the job of your national meteorological agency.
