WeatherNext

Learning Objectives

• Understand what WeatherNext predicts and why forecasting a cyclone's track and its intensity together is hard
• Identify the concrete accuracy gain the model delivers and how that translates into warning time
• Evaluate what open weights plus operational adoption at a national forecasting agency mean for the field

What Is WeatherNext?

WeatherNext is a machine-learning weather model from Google DeepMind and Google Research, built specifically for tropical cyclones — the storms known as hurricanes in the Atlantic and typhoons in the western Pacific. It was published in Nature on August 6, 2026.

The problem it addresses is a long-standing split in the field. Global atmospheric models are good at predicting where a storm will go but run at resolutions too coarse to capture how strong it will get. Specialized regional models capture intensity but do not see the whole atmosphere. Forecasters have historically had to blend outputs from both and accept the seams. WeatherNext predicts track, intensity, and wind structure jointly from one model, which is what lets it improve all three at once instead of trading them off.

The Accuracy Gain

WeatherNext's headline result is roughly an extra day of predictive accuracy: its three-day forecasts match what the previous generation of models achieved at two days. Google frames that as approximately a decade of conventional meteorological progress, delivered at once.

The resolution line is the counterintuitive one. WeatherNext runs on a 28-kilometer grid — around 100 times coarser than traditional models — and still beats them. That is the same lesson AI forecasting has taught repeatedly: a learned model does not need to resolve the physics explicitly if it has learned the patterns those physics produce.

A day of extra warning is not an abstract benchmark. It is the difference between ordering an evacuation with time to complete it and ordering one that traps people on the road.

Operational Use

WeatherNext is not a research demonstration waiting for adoption. It is already in operational use at:

• The US National Hurricane Center — the agency that issues official Atlantic and eastern Pacific forecasts and warnings
• The UK Met Office
• The Cooperative Institute for Research in the Atmosphere, a NOAA-affiliated research partner

Google cites Hurricane Melissa in 2025 as a validation case, where the model predicted both the rapid intensification and the Jamaica landfall in time to support advance warnings.

Pricing

There is no commercial tier. The models are published research artifacts, so the practical cost of running WeatherNext is whatever compute you put under it — which, at under a minute per forecast on a single accelerator, is modest by frontier-model standards.

Strengths

• A genuine, quantified accuracy gain — more than 24 hours of additional lead time on track, intensity, and wind structure, peer-reviewed in Nature rather than announced by press release
• Joint prediction — track, intensity, and structure come from one model, capturing the physical coupling that split models discard
• Cheap to run — a forecast in under a minute on a single tensor processing unit, against hours of supercomputer time
• Openly available — code and weights published, so the result is reproducible and deployable outside Google
• Already operational — in use at the National Hurricane Center, the UK Met Office, and CIRA, not pending evaluation

Limitations and Considerations

• Cyclones specifically — this is a specialized model for tropical cyclones, not a general-purpose forecasting system; WeatherNext 2 covers broader forecasting, and other tools cover regional and short-range needs
• Coarse grid — the 28-kilometer resolution works well for storm-scale prediction but is not the right instrument for local, street-level impact modeling
• Forecasts are not decisions — the model improves the input to an evacuation or preparedness decision; the official warning still comes from the National Hurricane Center, and operational meteorologists remain in the loop
• Trained on the past — like any learned model, it has seen historical storms. A genuinely unprecedented storm is exactly the case where a pattern-learned model has the least support

Related Tools

• WeatherMesh — WindBorne's AI forecasting model, paired with its own weather-balloon constellation
• AlphaFold — the earlier DeepMind result that set the template of open scientific models with broad adoption

Key Takeaways

• WeatherNext is Google DeepMind's cyclone forecasting model, published in Nature on August 6, 2026
• It predicts track, intensity, and wind structure jointly, gaining more than a full day of lead time — three-day forecasts now match what earlier models managed at two days
• It runs on a 28-kilometer grid and produces a forecast in under one minute on a single tensor processing unit, despite being far coarser than traditional physics-based models
• The code and model weights are openly published, and the National Hurricane Center, the UK Met Office, and CIRA already use it operationally
• The limits are scope, not quality: it is cyclone-specific, coarse-grained, and — like every learned model — least supported on genuinely unprecedented storms