Google 2015 logo (Wikimedia Commons, public domain)
WeatherNext 3 ingests live satellite data to produce global forecasts every hour, with precipitation accuracy improving up to 50% over previous models.
Weather forecasting has operated on roughly the same clock for decades. Traditional numerical weather prediction models ingest data every six hours, crunch massive physics simulations, and spit out results that are already slightly stale by the time anyone reads them.
On September 3, 2026, Google DeepMind and Google Research announced WeatherNext 3, an AI-powered global weather model that generates forecasts every single hour using live geostationary satellite observations. It’s the first global model to achieve true hourly update cycles, and early benchmarks suggest it’s not just faster but meaningfully more accurate than what came before it.
What WeatherNext 3 actually does differently #
The core innovation is deceptively simple in concept: instead of waiting for batched observational data to arrive every six hours, WeatherNext 3 trains directly on real-time satellite feeds. That means it can initialize a new forecast every hour, giving meteorologists and the industries that depend on them something much closer to a live picture of the atmosphere.
Resolution tells a similar story. The model delivers spatial granularity of up to 5 km for station-calibrated surface variables like temperature and moisture, 10 km for other surface-level measurements, and 25 km for atmospheric variables. That’s roughly five times sharper than its predecessor, WeatherNext 2, which launched in November 2025.
For precipitation, the improvements are even more dramatic. WeatherNext 3’s rain and snow forecasts are up to 50% more accurate than previous models for lead times of a day or longer, as measured by Brier score and CRPS against IMERG satellite precipitation observations. The model generates 64-member ensemble forecasts. Select forecast cycles extend out to a 15-day horizon, while the hourly interim runs cover a 48-hour window.
From GraphCast to WeatherNext 3: the lineage #
Google DeepMind has been methodically building its weather AI stack since 2023, when it introduced GraphCast, a model that demonstrated machine learning could match or exceed traditional physics-based forecasting for medium-range predictions. GenCast followed in 2024, adding probabilistic ensemble capabilities. WeatherNext 2 arrived in November 2025. WeatherNext 3 represents the next logical step: making the entire pipeline real-time by plugging directly into satellite observation streams rather than relying on pre-processed analysis fields that introduce latency.
Brightband’s Operational WeatherBench, an independent evaluation framework, has ranked WeatherNext 3 as the most advanced and accurate global weather model currently available.
The model also shows particular strength during extreme weather events. Cyclone track and intensity predictions have been a specific area of emphasis.
Clean energy and commercial applications #
WeatherNext 3 explicitly targets wind and sunlight variables, making it directly useful for renewable energy operations. Google is integrating the model’s outputs into Search, Maps, and Earth Engine. For developers and researchers, the data is accessible through BigQuery on Google Cloud.
The 50% improvement in precipitation accuracy at day-ahead timescales is particularly relevant for agriculture, since rain timing matters almost as much as rain volume for agricultural planning.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our