Google’s latest weather-model push rests on a simple premise: forecasts improve when the model sees the atmosphere closer to real time. WeatherNext 3, now feeding forecasts in Search, Maps and Gemini, shifts Google’s AI system from six-hourly updates toward hourly forecasts by ingesting satellite observations.¹

The company says that change delivers a global picture five times sharper than WeatherNext 2, with some variables mapped at up to 5-kilometre resolution rather than a 25-kilometre grid. Google research engineer Samier Merchant said the key was access to “fresher and richer observational data sets.”¹ Google says precipitation forecasts can be up to 50% more accurate at least a day ahead, particularly for rapidly evolving rain and snow systems and in regions with fewer ground gauges.¹

DeepMind chief Demis Hassabis amplified the rollout, describing WeatherNext 3 as a “major breakthrough” that learns from “real-world, real-time observations” to produce more localised predictions faster.Demis Hassabis (demishassabis)² Google is also pitching the model as an energy tool, including forecasts of wind speed at turbine height, as the company stresses the value of renewable generation amid rising data-centre demand.¹

But a closer reading of the technical results tempers the victory lap. An independent assessment found roughly a 5% improvement in upper-atmosphere accuracy over WeatherNext 2—equivalent, Google says, to about six additional hours of reliable lead time—and up to a 30% gain in location-specific surface-temperature accuracy.³

The same assessment identified an awkward early-forecast exception: for several variables, WeatherNext 3 initially performed worse than comparison models before moving ahead later in its 15-day outlook. It also flagged hexagonal-looking precipitation patterns and global temperature ensembles that can drift too warm or cool—oddities without an explanation in the paper.³

That is the central tension. AI can run far more cheaply and frequently than traditional numerical models, but it remains largely a pattern-learning black box. Google itself still trains WeatherNext 3 on physics-based-model data, while forecasters typically compare several models before issuing warnings.¹