Google DeepMind's WeatherNext 3 forecasts weather hourly at 5km

Google DeepMind and Google Research have released WeatherNext 3, an AI weather model built to update global forecasts every hour using satellite imagery, down to a resolution of 5 kilometers for temperature and humidity.

The model was announced on September 4, 2026, positioning it as the latest step in Google's push to replace or augment traditional numerical weather prediction with machine learning systems trained on atmospheric data.

What sets WeatherNext 3 apart from its predecessors is the combination of frequency and granularity. Rather than issuing forecasts on the multi-hour or daily cycles typical of legacy systems, it refreshes localized predictions hourly, and Google says those predictions are five times sharper than earlier WeatherNext versions.

That jump in detail matters for use cases where coarse regional forecasts fall short: flash flood warnings, wildfire risk assessment, aviation routing, and agricultural planning all depend on knowing what conditions look like at a specific location within the next few hours, not just the general trend across a wide region.

Built for cloud-scale access

Google is distributing WeatherNext 3 through its existing cloud infrastructure rather than as a standalone product. Developers and researchers can query the model's outputs through BigQuery, pull geospatial layers via Earth Engine, or integrate it directly into applications built on Google Cloud.

That distribution choice matters as much as the model itself. By embedding WeatherNext 3 into tools already used for climate analytics and satellite data processing, Google lowers the barrier for meteorological agencies, logistics companies, and disaster-response teams to plug high-resolution forecasts into existing pipelines without building new infrastructure.

Satellite imagery serves as the model's primary input, letting it generate forecasts without relying solely on ground-based sensor networks, which are sparse or absent across large parts of the ocean and less-developed regions. That approach could extend reliable short-term forecasting to areas historically underserved by conventional weather models.

No pricing details or API rate limits were disclosed alongside the announcement. It also remains unclear how WeatherNext 3's accuracy compares directly against national forecasting agencies' own high-resolution models, a benchmark that will likely determine how quickly agencies and enterprises adopt it over incumbent systems.