The Google WeatherNext 3 AI weather model went live across Search, Gemini, Maps and Cloud on Thursday, 3 September 2026, refreshing global forecasts every hour instead of every six.
The system arrived as the most advanced and accurate global weather model the company has built, according to Google, which leaned on independent live evaluations by a third party called Brightband.
The leap was less about scale than about where the model learned its lessons.
How the Google WeatherNext 3 AI weather model dropped the six-hour lag
Numerical weather prediction was the nerdy phrase at the heart of the old approach, meaning physics simulations run on supercomputers. Most AI forecasters, WeatherNext 2 included, trained on the output of those runs, and that output carried a six-hour lag.
WeatherNext 3 skipped the middleman and trained directly on live, raw satellite observations, the mosaic of images and readings beamed down as the atmosphere actually behaves.
Removing the lag let the model produce a fresh forecast every hour rather than every six.
Resolution improved just as sharply. The model produced forecasts for key surface variables such as temperature and moisture at 5km, other surface variables at 10km, and atmospheric variables like wind speed at 25km, against a flat 25km for WeatherNext 2.
Overall that made it roughly five times sharper than its predecessor.
What the Google WeatherNext 3 AI weather model means for rain forecasts
Rain was where the upgrade showed up most plainly. The company said anyone planning a day or more ahead would see precipitation forecasts up to 50% more accurate, with the biggest gains in regions where forecasting had long been least reliable.
Continuous Ranked Probability Score, or CRPS, was the yardstick in lab testing, a measure of how closely a probability forecast matched what the sky actually did.
On that scale, the model improved by up to 60% against NASA’s satellite-based IMERG rainfall data, 30% against radar-based MRMS data and 10% against rain-gauge measurements.
Wind, sunlight and the regions left behind by traditional forecasting
Renewable energy planners got their own layer. The model forecast wind speeds at 100m, roughly the height of a turbine hub, alongside high-resolution cloud cover and solar radiation figures that let solar farms estimate how much sunlight would land on their panels.
The satellite-first approach mattered most where regional supercomputer models had been too costly to run.
Google flagged it as particularly vital across Latin America, Africa and Asia-Pacific, regions long stuck with coarse forecasts because the traditional modelling bill was prohibitive.
Access was not limited to the apps. Developers and researchers could query the raw forecast data through BigQuery or Earth Engine, or bulk-download it from Google Cloud Storage, alongside the Maps Platform Weather API already wired into the new model.
WeatherNext 3 will keep filtering through the products it powers, from Search results to Maps and the Gemini app.
South African users sit inside the band of regions Google singled out as underserved, which matters in a week when Swisher Post reported a level 5 storm warning over Gauteng and the Northern Cape.
Better rain forecasting lands where it is needed most.
