Google’s WeatherNext 3 Redraws the Forecast Every Hour

The forecast you checked at breakfast was probably built from a picture of the atmosphere taken several hours before you looked at it. That lag is baked into how traditional forecasting works, and it is the thing Google DeepMind went after with WeatherNext 3, announced Thursday and already wired into Search, the Gemini app and Google Maps.

The big change is cadence. The previous version put out a new forecast every six hours. This one does it every hour, pulling in live geostationary satellite imagery as it arrives instead of waiting on the usual analysis cycle. Google calls the result "roughly five times sharper than our previous model."

What "sharper" actually buys you

Sharper means smaller boxes on the map. WeatherNext 3 predicts 2-meter temperature and dew point at roughly 5km resolution, other surface variables like wind, pressure, cloud cover and hourly precipitation at about 10km, and the upper-atmosphere fields at about 25km across 13 pressure levels. The previous model worked at 25km for everything.

Five kilometers is about the difference between "it will rain in your metro area" and "it will rain on your side of town." That is the scale where a forecast starts to be worth acting on, and it is exactly where consumer weather apps have always felt vague.

It skips the physics simulation

Traditional numerical weather prediction solves the physics of the atmosphere on a supercomputer. That is accurate, and it is slow. WeatherNext 3 learns from data instead: ERA5 reanalysis, NASA's IMERG satellite precipitation retrievals, raw weather station measurements, and hourly satellite mosaics. Google says training directly on station observations, rather than only on smoothed reanalysis, is part of why the local numbers hold up.

The full runs go out 15 days with 64 ensemble members at the four standard synoptic cycles, with shorter hourly runs covering the next 48 hours. Ensemble members are the model running the same forecast many times with small variations, which is how you get "40 percent chance of rain" instead of a single guess.

Google's numbers, and who checked them

Google claims a precipitation CRPS improvement of up to 60 percent against IMERG, 30 percent against the MRMS radar product and 10 percent against rain gauge measurements at early lead times, plus up to 50 percent more accurate precipitation forecasts a day or more ahead. CRPS scores a probabilistic forecast against what actually happened, and lower is better. These are Google's own measurements, run by the team that built the model.

Google also says independent live evaluations from Brightband rank WeatherNext 3 as the most accurate global weather model to date. Brightband is a weather AI company that runs Operational WeatherBench, a live leaderboard that scores physics-based and AI models against real observations as each forecast cycle completes, and its benchmarks page does list WeatherNext 3 as the new leader. So the claim checks out at the source. It measures routine forecast skill, though. The storms people evacuate for are a separate test, which Brightband runs as a different benchmark entirely.

The extreme weather question

AI forecasting has a record to answer for here. A study published in Science Advances in April, led by Zhongwei Zhang at the Karlsruhe Institute of Technology, checked GraphCast, Pangu-Weather and FuXi against ECMWF's physics-based HRES using roughly 160,000 heat records, 33,000 cold records and 53,000 wind records from 2020. The AI models underestimated both how often record-breaking events happened and how extreme they got. Erich Fischer, one of the researchers quoted on the findings, called it a warning shot against swapping out physics models too quickly.

Hurricanes show a similar split. A Rice University study published in March in the Journal of Geophysical Research: Atmospheres ran Pangu-Weather and Aurora against about 200 tropical cyclones from 2020 through 2025. Both tracked storm paths well and got shakier on intensity and structure. Lead author Avantika Gori put it plainly: "windfields can look realistic while still violating key aspects of atmospheric physics."

Neither study tested WeatherNext 3, and Google says this model does better in regions where training data is thin. But the pattern repeats across model families, so the burden of proof sits with the new one until forecasters have run it through a full season.

Where you'll run into it

Most people will meet WeatherNext 3 without noticing, through Search, Gemini and Maps. Developers get it through the Google Maps Platform Weather API, Google Earth Engine, BigQuery queries and bulk downloads from Google Cloud Storage, with no model to stand up yourself. The weights are not open source, and on-demand custom inference still runs the older WeatherNext 2.

Google's own fine print tells you to go to your local meteorological agency or national weather service for official forecasts and severe weather warnings, which is the right instinct. Use this one to decide whether you need a jacket. When a storm is actually coming, keep listening to the National Weather Service.