Agent Horizon

Google DeepMind and Google Research introduced WeatherNext 3, a new global weather forecasting model that generates predictions every hour instead of the usual six, drawing directly on live satellite data rather than waiting on slower government-produced datasets. Announced via Google’s official blog post, the model is already rolling into Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine, with developer access also available through BigQuery and Cloud Storage.

The headline technical shift is training methodology: WeatherNext 3 learns directly from raw satellite observations and weather station data instead of relying primarily on the outputs of traditional numerical weather prediction systems, which carry a six-hour lag before a forecast can even be produced. That change lets the model refresh continuously and resolve surface variables like temperature and moisture down to a 5-kilometer grid, a five-fold sharpening compared to its predecessor’s 25-kilometer resolution. According to independent live evaluations from the benchmarking startup Brightband, the model currently sits atop the Operational WeatherBench leaderboard, ahead of physics-based systems from ECMWF and NOAA as well as deep-learning competitors from other tech companies.

This is a genuinely useful, unglamorous kind of AI progress — the sort that rarely gets a keynote moment but quietly touches how billions of people plan their day. Faster, sharper precipitation forecasting has real value beyond convenience: DeepMind is specifically targeting wind and solar operators with new 100-meter-altitude wind speed and solar radiation predictions, aimed at helping grid operators match renewable output to demand more precisely. It’s also a good example of AI research translating into shipped infrastructure rather than staying a benchmark curiosity, since the model is live in consumer products on day one rather than being a research preview.

The caveats are worth stating plainly. Google is citing its own selected metrics alongside a benchmark run by a smaller independent startup, so some of the most eye-catching numbers — up to 60% improvements in certain precipitation scores — deserve a healthy dose of skepticism until more independent, longer-run comparisons accumulate. And despite the accuracy gains, Google is careful to still direct people to official government agencies like the National Weather Service for actual severe weather warnings, a sensible boundary for a system that, however impressive, is still a statistical model rather than a substitute for emergency meteorological infrastructure.