Google’s DeepMind team has introduced GenCast, a new AI model for weather prediction that it says outperforms a leading operational forecast system. The claim is based on a comparison with the European Centre for Medium-Range Weather Forecasts’ ENS, described in the source as apparently the world’s top operational forecasting system.
The announcement matters because weather forecasting is not just about saying whether it may rain. Forecasts guide planning, risk assessment, travel decisions, public information, and research. DeepMind is presenting GenCast as part of a broader move to bring AI-based weather models into more visible and useful places, including Google Search and Maps.
What makes GenCast different
GenCast is designed around uncertainty. Instead of producing only one best guess about future conditions, the model generates many possible paths that the weather could take.
DeepMind contrasted GenCast with its previous weather model, which it described as deterministic. That older approach produced a single estimate of future weather. GenCast takes a different route by building an ensemble of forecasts.
According to the DeepMind team, GenCast includes 50 or more predictions. Each prediction represents a possible weather trajectory. Taken together, those predictions form a probability-based view of what could happen next.
That distinction is important in plain language. A single forecast can be useful, but weather is complex. A set of possible outcomes can help show where the forecast is more settled and where it is less certain.
How DeepMind measured it against ENS
DeepMind researchers said in a paper published in Nature that GenCast outperforms the European Centre for Medium-Range Weather Forecasts’ ENS. The comparison was built around a specific training and testing setup.
The team said GenCast was trained on weather data up to 2018. It then compared GenCast forecasts for 2019 with ENS forecasts for the same period.
DeepMind said GenCast was more accurate 97.2 percent of the time in that comparison. The source does not give further detail on every metric behind that result, so the key point is the one DeepMind reported: in its test against ENS, GenCast came out ahead most of the time.
The comparison also shows why AI weather models are being watched closely. Forecasting systems are judged by performance across many situations, not by one isolated example. DeepMind’s stated result positions GenCast as a serious entrant in operational weather prediction discussions.
Why an ensemble forecast matters
The central idea behind GenCast is not only that it predicts weather, but that it represents a range of possible futures. That can make a forecast more useful when the future is not clear-cut.
A probability distribution gives users a structured way to think about uncertainty. In everyday terms, it means the model is not forced to pretend there is only one possible answer. It can show that several outcomes remain possible.
That approach fits the nature of weather itself. Weather systems can shift, and small differences can matter. A forecast model that keeps multiple scenarios in view can provide more context than a single path.
For researchers, that kind of output can also be valuable because it can feed into other models and studies. DeepMind said it plans to release real-time and historical forecasts from GenCast, which anyone can use in their own research and models.
Where Google plans to use GenCast
Google says GenCast is part of its suite of AI-based weather models. The company is also starting to incorporate those models into Google Search and Maps.
That could make AI-generated weather forecasting more visible to everyday users. Search and Maps are places where people already look for immediate, practical information. Adding AI-based weather models there could connect the technology to routine decisions.
The source does not say exactly how GenCast will appear inside those products, or how much of the user experience will change. What is clear is that Google is treating AI weather prediction as more than a research project.
The planned release of GenCast forecasts also points to a wider use case. By making real-time and historical forecasts available, DeepMind is giving researchers and model builders a way to work with the system’s outputs directly.
What to watch next
GenCast’s headline result is strong: DeepMind says it beat ENS in its 2019 forecast comparison 97.2 percent of the time. But the broader significance is the direction of travel. AI weather models are moving from laboratory results toward public-facing products and shared forecast resources.
Several facts stand out from the announcement:
- GenCast was unveiled by Google’s DeepMind team.
- The model uses an ensemble of 50 or more predictions.
- DeepMind compared GenCast with the European Centre for Medium-Range Weather Forecasts’ ENS.
- GenCast was trained on weather data up to 2018 and tested against forecasts for 2019.
- Google plans to bring AI-based weather models into Google Search and Maps.
- DeepMind plans to release real-time and historical GenCast forecasts for use in research and models.
The practical takeaway is simple. GenCast is Google DeepMind’s attempt to make weather prediction more probabilistic, more accurate in its own benchmark, and more useful beyond a single research paper. If the released forecasts become widely used, the model could become an important tool for researchers working with weather data and for Google users who rely on forecast information in daily products.