The Weather Forecast Is Now an AI Model

For most of the last decade, the question about AI weather models was whether they could match the physics-based systems run by national agencies. WeatherNext 3, released by Google DeepMind and Google Research on 3 September 2026, changes the question. It is the first operational global model to refresh every hour, it learns directly from satellite and station observations rather than from another model's output, and it is now the forecast that billions of people see in Google Search, Maps and Gemini.

What changed

The launch post describes three shifts. Resolution: temperature and moisture at 5 km, other surface variables at 10 km, atmospheric variables at 25 km, roughly five times sharper than WeatherNext 2. Cadence: the model ingests a mosaic of live geostationary satellite data and issues a new forecast every hour, whereas models trained on numerical weather prediction carry what Google calls a six-hour data lag. Skill: Google reports precipitation improvements in continuous ranked probability score of up to 60% against NASA's IMERG product, 30% against MRMS radar and 10% against rain gauges at early lead times.

The accompanying paper on arXiv gives the architecture: a functional generative network mesh transformer, with latent size increased from 768 to 1,024 and mesh depth from 24 to 32 layers over the previous model, trained through a curriculum from one degree to 0.1 degree and using ERA5 reanalysis back to 1959. The live satellite feed uses 11 spectral channels from the GOES, Meteosat and Himawari satellites, regridded to 0.1 degree, with an archive back to 2016.

Against the incumbents

The paper reports a quasi-real-time comparison from 1 July to 11 August 2026 in which the production model, trained through 30 June, outperformed the European Centre for Medium-Range Weather Forecasts' AIFS ENS v2 on what it calls a vast majority of analysis metrics, while cautioning that six weeks is a small sample. The independent check is Brightband's Operational WeatherBench, which tracks live forecasts: on 850 hPa temperature error over days one to seven, WN3 ranks first, winning 18 of the last 30 global cycles, with WeatherNext 2 second on 11 and ECMWF's AIFS ENS third with one. Microsoft's Aurora, ECMWF's IFS ensemble, NVIDIA's Atlas and NOAA's GEFS and GFS won none in that window. WeatherNext 2 still leads over Europe.

The energy variables

WeatherNext 3 adds 100-metre wind speed, roughly turbine hub height, along with high-resolution cloud cover and solar radiation. Google DeepMind research scientist Ferran Alet told The Verge that making renewables 'a very appealing opportunity is very important for us' as Google's and humanity's energy needs rise. An experimental 0.05-degree dataset is available on Earth Engine under a CC BY 4.0 licence, with a warning that it is not validated for real-world use, and Google's blog directs users to national meteorological agencies for official warnings.

Who benefits, who is at risk

Beneficiaries: wind and solar operators, grid schedulers and energy traders with access to hourly, turbine-height forecasts; Google, which now owns a forecast advantage across its consumer products. At risk: commercial forecast vendors reselling agency output, and national agencies whose public role now depends on independent benchmarks to demonstrate parity.