Google's WeatherNext 3 Sharply Upgrades AI Weather Forecasting
Google announced on September 3, 2026, that it is rolling out WeatherNext 3, the latest version of its AI weather forecasting model, with what the company describes as substantially improved accuracy in predicting rain and snowfall at least one day in advance. Google claims precipitation forecasts are up to 50 percent more accurate than its previous model when looking at least a day ahead. The Verge
The headline technical upgrade is resolution — both in space and in time. WeatherNext 2, the previous model, divided the globe into a grid of 25-kilometer squares and produced a new forecast every six hours. WeatherNext 3 can visualize variables like temperature and moisture at up to 5-kilometer resolution and generates a new forecast every hour, using live satellite data fed into each prediction cycle. Think of it as switching from a camera that takes a wide, blurry photo twice a day to one that snaps a sharper picture every hour. Samier Merchant, a research engineer at Google Research, said the model learns from real-time weather observations and what he described as "fresher and richer" observational data sets. The Verge
WeatherNext 3 also introduces predictions tailored to renewable energy generation, including wind speed at 100 meters — roughly the height at which a utility-scale wind turbine's blades rotate. Google identifies agriculture, renewable energy, and daily planning as the areas where the model offers significant improvements. Google Research Blog
The model is not a clean break from traditional forecasting. WeatherNext 3 is still trained on data from physics-based weather models — the kind that use mathematical equations to simulate how the atmosphere behaves — and is expected to work alongside traditional numerical weather prediction rather than replace it. This fits the trajectory of the WeatherNext family, which Google DeepMind and Google Research have built as a set of global, medium-range atmospheric models using machine learning. The Verge
Google has built a track record with the WeatherNext line in the months leading up to this release. The WeatherNext 2 model, introduced in November 2025, generates hundreds of possible weather scenarios in under a minute on a single TPU (Google's custom AI chip) and delivers forecasts up to 15 days ahead. In a paper published in Nature, Google researchers showed that the WeatherNext AI model achieved state-of-the-art accuracy in predicting a cyclone's track, intensity, and wind structure, and enabled cyclone forecasts that can give an extra day of warning. The model was used by the US National Hurricane Center to predict Hurricane Melissa's Category 5 landfall in Jamaica. Google DeepMind Blog
Google has collaborated with the US National Hurricane Center and agencies in Asia to improve forecasting using its AI models, and has made WeatherNext models accessible via Google Cloud's Vertex AI Model Garden for customization and deployment in research and industry settings. The company also launched Weather Lab, an interactive site for sharing its AI weather models and experimental cyclone predictions. DeepMind has open-sourced the WeatherNext AI model. The Verge
On the consumer side, WeatherNext 3 is incorporated into Google Search, Maps, Gemini, and other Google products, meaning the improved resolution and hourly update cycle will feed directly into services that hundreds of millions of users consult daily. The Verge
The jump from a 25-kilometer, six-hour cadence to a 5-kilometer, hourly cadence is where WeatherNext 3 earns its version number. For a model running globally, that is not a linear increase in computing power; it is closer to a 25-fold increase in spatial grid cells and a sixfold increase in temporal frequency, before accounting for the additional observational data pipelines feeding each cycle. The fact that Google is folding this into consumer-facing products like Search and Maps suggests the inference cost per forecast has dropped to a level where serving it at scale is economically viable, though Google has not disclosed specific cost or latency figures.
The renewable energy angle is worth noting. Wind speed at turbine hub height is a variable that matters directly for grid operators and energy traders, not just for weather enthusiasts. If WeatherNext 3's claimed 50 percent improvement in precipitation accuracy at one-day lead times extends even partially to wind and solar generation forecasting, the model becomes a planning tool for an industry where small forecast errors translate into real dispatch costs — the cost of routing power plants to meet expected demand.
Google's decision to keep WeatherNext 3 trained on physics-based model data, rather than training purely on raw observations, fits the broader consensus in AI weather modeling: machine learning systems still benefit from the physical constraints encoded in numerical weather prediction, and the hybrid approach outperforms either method alone for now. The open-sourcing of WeatherNext and its availability through Vertex AI also positions Google as an infrastructure provider for meteorological research, not just a consumer of forecasts for its own products.
What remains unaddressed in the announcement is how WeatherNext 3 performs on extreme and rare events, where training data is inherently sparse, and how the hourly update cycle handles the delay between satellite observation ingestion and forecast publication. The cyclone work with the National Hurricane Center and the Nature paper suggest the model family performs well on tropical systems, but operational validation at the new resolution and cadence will take time. For now, the model is live, and the fivefold resolution increase and hourly cadence set a new baseline for AI-driven global weather forecasting that competing efforts will need to match.
