Google's new weather model refreshes hourly and lands inside Search, Maps and the Gemini app

The numbers, and where each one comes from

• Every hour: how often a new global forecast is produced, from Google’s announcement of 3 September 2026.
• 5, 10 and 25 kilometres: the grids for key surface variables, other surface variables and upper-air variables, against WeatherNext 2 on a 25-kilometre grid in 6-hour steps.
• Up to 50% more accurate precipitation forecasts when planning a day or more ahead, Google’s own figure, with the largest gains where forecasts have been least reliable.
• About six hours: the lag in the numerical weather prediction data most AI weather models are trained on, which this one replaces with live satellite imagery.
• $0, $4.99, $19.99 a month for the Gemini tiers a person can buy, as our table recorded them on 5 September 2026, none of which this release changes.

Why the training data is the story

The resolution numbers are the easy headline. The change underneath them is what the model learns from: most AI forecasting models are trained on the output of supercomputer physics runs, which arrive hours after the weather they describe. Training on live satellite imagery and on sparse weather-station readings instead means the model is fitted to observations rather than to somebody else’s simulation of them.

That is also the honest caveat. A model trained on observations inherits the gaps in observation coverage, and the places with the thinnest station networks are exactly the places Google says improve most. Both things can be true, and only a year of live scoring will separate them.

What we will watch

This site tracks what Gemini costs, not what the weather does, so the row that matters here is the plan ladder above: a capability of this size arriving with no price attached is a pattern worth dating, because the same company published a promotional API price last week with an end date printed on it. When a free capability starts carrying a fee, our plans table is where that will show up.