Google WeatherNext 3, the company’s latest AI forecasting model, began rolling out on September 3 with a new input: recent satellite observations. Google says the change will produce global forecasts every hour and add useful detail for rain and snow, particularly where weather stations on the ground are scarce.
The practical change is less mystical than the AI branding suggests. Conventional forecasting systems simulate atmospheric physics with large numerical models. Google’s system learns weather patterns from data, then takes in fresh satellite observations to update its estimate of current conditions. That faster refresh is intended to help with precipitation systems that can move and change quickly.
What changed in Google WeatherNext 3?
According to The Verge, WeatherNext 2 generated forecasts on a 25-kilometer grid at six-hour intervals. WeatherNext 3 can represent some variables, including temperature and moisture, at up to 5-kilometer resolution and creates a forecast each hour from the latest satellite data.
Those are three distinct measures that tend to get tossed into one marketing bucket: how often a model updates, how detailed its map is, and whether its forecast is right. Hourly updates and a finer grid make the output more timely and more granular. They do not, by themselves, establish higher accuracy.
Google says the new model can make precipitation forecasts up to 50% more accurate at least a day ahead. The company expects the largest gains in areas with fewer rain gauges, mainly outside the US and Europe, where satellite observations can help fill holes in surface measurements. That performance figure is Google’s claim, reported by The Verge, rather than an independently documented assessment in the available reporting.
Where will WeatherNext 3 show up?
Google says it is incorporating WeatherNext 3 into Search, Maps, Gemini and other products. TechCrunch also reported that Google intends to make the model available through its cloud platforms. Rollout details and the extent of product integration were not specified in the reports.
The company also built outputs for energy forecasting, including wind-speed predictions at 100 meters, roughly turbine height. That could be useful for estimating wind generation, though a forecast model does not settle operational decisions on its own.
Will AI weather forecasts replace regular weather models?
No such replacement is established here. Weather agencies use several forecast systems when issuing warnings, and The Verge reports that WeatherNext 3 is trained using data from physics-based models. The older systems remain part of the machinery that supplies the data and comparative forecasts.
Google has other AI-weather projects, including GenCast, a separate ensemble model. In a 2024 post about a Nature-published study, Google DeepMind said GenCast generated multiple possible weather scenarios to express uncertainty. That work is background, not proof that WeatherNext 3 meets the same performance bar.
For people checking whether to carry an umbrella, the advertised benefit is more frequent, finer-grained forecast information. For emergency warnings, the relevant forecast remains the one issued by the responsible weather agency after it weighs the available models.
This story draws on original reporting from The Verge.
