Google’s New Forecast Model Could Change Storm Planning

The most important change in weather forecasting right now isn't that AI can predict rain or sun a little better. It's that forecasting is quietly turning into a business scenario-because for storms, heatwaves and energy planning, the question is rarely "what's the most likely outcome?" and more often "what's the worst one we still need to be ready for?"

Google DeepMind and Google Research introduced a new weather-forecasting AI model on November 17, 2025, called WeatherNext 2. It is being used in regular weather apps and services developed by Google. This shows DeepMind believes these AI models are ready to be used as everyday tools, not just experiments.

Deepminds first version WeatherNext 1 was already a major step and now compared to that WeatherNext 2 is eight times faster and more accurate on 99.9% of forecast variables and lead times, while also upgrading temporal resolution to hour-by-hour predictions.

WeatherNext 2 can forecast the weather on an hour-by-hour basis as far out as 15 days. According to Google, hundreds of different possible forecast outcomes can be created in less than a minute using one of its TPU chips. Traditional supercomputers generally take hours to accomplish the same task.

The big improvement comes from a new kind of model called a Functional Generative Network(FGN). Older AI weather systems tried to learn everything at once, but FGN works differently. First, it learns to make good predictions for each weather factor on its own, such as temperature, wind, or rain. It does this by using a training method that rewards accurate forecasts.

Then FGN learns how those factors fit together in real life, such as how temperature and pressure change together across different places. It does this by adding a small amount of carefully controlled randomness inside the model. This allows the model to produce many different possible weather scenarios fast while still keeping them realistic and consistent with real-world physics.

That move matters because probabilistic forecasting is where AI has lagged behind its deterministic wins. DeepMind's earlier GenCast model was the first ML ensemble to clearly beat ECMWF's ENS system. FGN is designed as the next step: faster, easier to scale and stronger on calibration and extremes.

And you can already see why Google wants this stuff out in the field. In March this year, experimental tropical-cyclone forecasting began from DeepMind through Weather Lab and pre-release readiness of the product, WeatherNext 2. Those cyclone runs were not merely pretty maps; they were constructed to aid agencies and planners in assessing multiple tracks and intensities days ahead of what was previously possible.

This is where it starts to be real scale on the consumer side. WeatherNext 2 is already driving improved forecasts in Google Search, Gemini responses and the Pixel Weather app; it’s also being incorporated into Google Maps Platform's Weather API with full Maps surfaces to come over the next few weeks.

Google is also opening the pipes for developers and enterprise teams: WeatherNext 2 forecast data is available in Google Earth Engine and BigQuery, while Google Cloud is operating a Vertex AI early-access program that allows organizations to run custom inference on top of the model family.

Stepping back, WeatherNext 2 hits in a rising competition wave, with the likes of ECMWF, Nvidia, Huawei, and more pushing their own ML-forecasting stacks; the gap between research paper and operational use is shrinking fast. What separates this release is less the raw idea of "AI weather" but more the institutional signal: Google is committing its flagship products to an ML ensemble as the backbone forecast. 

If WeatherNext 1 was proof that AI could forecast well, WeatherNext 2 is Google saying it's ready to be the default weather brain behind consumer apps and industrial decision systems alike. And with scenario forecasting now running fast enough, the bigger shift is cultural: we're moving from "What will happen?" to "What could happen?” and "how soon do you need to act?"