Google DeepMind and Google Research released WeatherNext 3 on Sept. 3, a global weather model that produces a new forecast every hour instead of every six and resolves surface conditions down to 5 kilometers. The model is now powering weather features inside Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API and Google Earth Engine, with the raw forecast data queryable in BigQuery and Earth Engine or downloadable in bulk from Cloud Storage. That distribution footprint, more than any single benchmark, will decide whether WeatherNext 3 changes what shows up on a phone screen before a commute.
Weather forecasting is one of the few domains where a learned model has to beat a physics simulation on the simulation’s own turf: fluid dynamics, thermodynamics, decades of validated numerical weather prediction. WeatherNext 3 breaks from Google’s prior approach by training directly on raw geostationary satellite mosaics and weather-station readings, rather than on the output of those physics-based systems, which the company says carry a six-hour lag that biases fast-moving variables like rainfall and surface temperature.
The number worth tracking here is not accuracy on its own. It is accuracy per unit of compute, because that ratio decides whether a national meteorological agency, not just Google, can afford to rerun a model at this resolution every hour rather than every six. A learned model that matches an NWP system’s skill at a fraction of the supercomputer time is the real disruption to that agency’s budget. One that merely ties on accuracy while costing the same to run changes nothing operationally. Google has not published a compute-cost comparison against a physics-based system of equivalent resolution, an omission worth noting given how central the hourly-refresh claim is to the launch.
Google reports its own Continuous Ranked Probability Score, a standard measure of probabilistic forecast accuracy, improved by as much as 60 percent versus NASA’s IMERG satellite rainfall dataset, 30 percent versus the MRMS radar network, and roughly 10 percent versus ground rain gauges when lead times are short. Google also says forecasts made a day or more in advance are as much as 50 percent more accurate than WeatherNext 2’s, with the biggest gains in regions that historically had the thinnest forecasting coverage. Every one of those figures is Google comparing WeatherNext 3 against its own prior model and against reference datasets it selected; none of it is an independent skill-score audit, and Google’s own post says so only implicitly.
The one external check Google cites is Brightband, a nonprofit that runs continuously updated leaderboards comparing weather models against verified observations. Google says WeatherNext 3 currently ranks as the most accurate global model on that board, a genuinely independent signal. It is still one leaderboard measuring one slice of skill, and Google did not disclose where WeatherNext 3 lands against non-Google entrants at lead times beyond the medium range, or how its ranking has moved week to week.
The launch’s most concrete addition is aimed at a narrower audience than consumer weather apps. WeatherNext 3 now forecasts 100-meter wind speeds at roughly turbine height, along with cloud cover and solar radiation, explicitly so grid operators can match renewable output to demand. For any team currently pulling forecasts from Google’s older WeatherNext 2 endpoint, an NWP-derived feed, or a competing satellite-trained model for energy planning, the hourly refresh and 5-kilometer resolution are worth a direct check against Brightband’s public leaderboard before switching, since that is the only figure in this launch that is not Google grading its own homework.
Google’s WeatherNext team detailed the model in a company blog post published Sept. 3, 2026.
