Google DeepMind's GenCast marks a major step for AI weather forecasting. Research published in Nature shows the system outperforming the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble system in 97.2 percent of test cases, while generating forecasts far more quickly.
Why the comparison matters
ECMWF's ensemble system is described as the world's top weather prediction system, so GenCast was measured against a demanding benchmark. The contrast is not only about accuracy. It is also about how much computing is needed to produce a usable forecast.
The ECMWF system needs several hours on a supercomputer with tens of thousands of processors. GenCast creates a 15-day forecast in just eight minutes using a single Google Cloud TPU v5 chip.
That speed matters because weather forecasting is most useful when it can be updated, compared, and acted on quickly. A faster system can make the same forecasting window more practical for teams that need to evaluate changing conditions.
GenCast forecasts many possible futures
GenCast does not produce only one answer. Like ECMWF's ensemble system, it creates more than 50 possible weather scenarios.
That approach is important because perfect weather forecasts are not possible. The value of an ensemble forecast is that it maps probability across multiple possible outcomes instead of presenting a single path as certain.
For decision-makers, that can be more useful than a simple yes-or-no forecast. Severe weather, energy planning, and disaster response all depend on understanding the range of likely outcomes, not just the most central prediction.
Severe weather is a core strength
The source research highlights GenCast's performance in severe weather forecasting. The system delivered more accurate forecasts for tropical cyclones, extreme temperatures, wind speeds, and air pressure.
One example is Typhoon Hagibis in October 2019. GenCast accurately mapped the storm's potential path seven days before it reached Japan, and its accuracy increased as the typhoon approached.
That kind of forecasting capability is especially relevant because the hardest weather events to plan for are often the ones where timing, direction, pressure, wind speed, and temperature extremes all matter at once. The article does not claim perfect prediction, but it does show stronger results across important severe-weather categories.
What GenCast is built on
GenCast was trained on 40 years of historical weather data through 2018, using ECMWF's ERA5 archive. It uses diffusion technology, the same broad family of methods associated with AI image generation systems, but adapted for weather prediction.
The model works at 0.25-degree resolution and can predict over 80 different weather variables. Those variables include conditions at ground level and in the atmosphere.
DeepMind is making the code and model details available to scientists. That open release is intended to help researchers, meteorologists, data scientists, renewable energy companies, and organizations focused on food security and disaster response improve weather forecasting.
Benefits, limits, and the wider race
Testing also indicates GenCast can predict wind power generation more precisely than current systems. That could help power companies better integrate wind energy into electrical grids.
The system still has limitations. Its resolution needs improvement to match ECMWF's recent upgrade to 0.1 degrees, and it requires more computing power than similar AI models.
GenCast follows DeepMind's earlier GraphCast system. The wider field is also moving quickly: the European Weather Center rolled out its own AI system built with Huawei technology, Nvidia recently demonstrated its own weather prediction advances, and Google's teams are developing additional AI forecasting systems as well.
The main takeaway is clear: AI weather forecasting is no longer only an experimental alternative to established systems. In GenCast's case, it has beaten a leading traditional ensemble system across a large majority of test cases, while producing 15-day forecasts in minutes rather than hours.