Weather forecasting has always depended on seeing the atmosphere again.
The question is how quickly a forecast system can absorb a new view of the planet and turn it into another prediction.
On September 3, Google DeepMind and Google Research introduced WeatherNext 3, an artificial-intelligence weather model that directly incorporates live geostationary satellite observations and can initialize a new global forecast every hour.
For selected surface variables, Google says the system reaches resolutions as fine as approximately 5 kilometers. Other surface and atmospheric variables are produced at coarser resolutions.
Google is also integrating WeatherNext 3 into Search, Gemini, Maps, Google Maps Platform, Cloud, and Earth Engine.
That combination makes the release more than another benchmark result.
A learned weather model is being connected more tightly to fresh observations, updating more frequently, resolving some fields more finely, and moving into services used for everyday decisions.
It does not mean artificial intelligence has solved weather forecasting.
A forecast can be hourly in two different ways
Weather products often describe forecasts in hourly increments.
That does not necessarily mean the underlying forecast itself was initialized an hour ago.
A model can produce hour-by-hour predictions from an atmospheric state that was estimated several hours earlier.
WeatherNext 3 changes that cadence.
Google says the system can initialize a new forecast every hour using live geostationary satellite input.
That distinction matters most when the atmosphere is changing quickly.
The fresher the starting state, the less time there is for the real atmosphere to diverge from the observations used to begin the forecast.
Whether that produces better decisions across every hazard and region remains an empirical question.
But hourly reinitialization is a meaningful operational change.
Learning directly from the sky
Weather prediction has long depended on enormous observation networks: satellites, weather stations, balloons, aircraft, radar, buoys, and other instruments.
Those observations are traditionally combined through sophisticated numerical analysis systems before many forecasting models begin their calculations.
WeatherNext 3 does not remove that meteorological infrastructure.
Instead, Google says the model directly incorporates live geostationary satellite mosaics as part of its initialization.
That allows a learned forecasting system to use a more immediate observational stream rather than depending only on slower upstream analysis products to describe the latest atmospheric state.
This is part of a larger change in artificial intelligence.
Learned systems are increasingly being trained not only on human-produced language, images, and code, but on measurements of the physical world itself.
In this case, the target is the atmosphere.
Higher resolution is not the same as greater truth
Google also increased the spatial resolution of WeatherNext 3.
Selected surface variables, including temperature and dew point, are forecast at approximately 5-kilometer resolution. Other surface variables are produced at approximately 10 kilometers, while atmospheric variables remain coarser.
Google describes the highest-resolution fields as roughly five times sharper than comparable WeatherNext 2 outputs.
More spatial detail can make forecasts more useful for local decisions.
It can also make errors look more precise.
Google’s own limitations documentation notes artifacts in some predicted fields, including stronger artifacts in some station and precipitation outputs.
A forecast drawn on a finer grid is therefore not automatically a better forecast.
Resolution, calibration, accuracy, uncertainty, and decision value are separate properties.
An independent scoreboard, with boundaries
Google describes WeatherNext 3 as its most advanced and accurate global weather AI model.
That claim is not supported only by Google’s internal evaluation.
Brightband independently tracks WeatherNext 3 on Operational WeatherBench, a live framework that compares artificial-intelligence and numerical weather-prediction systems across variables, regions, lead times, and models.
That external evaluation strengthens the evidence that WeatherNext 3 is competitive with leading forecasting systems on the variables Brightband currently scores.
It does not establish universal superiority.
Brightband’s methodology itself documents comparison limits.
In particular, WeatherNext 3 precipitation scores are currently omitted from the common leaderboard because its precipitation target and resolution do not allow a clean comparison under the existing methodology.
That matters because precipitation is one of the areas where Google emphasizes new capability.
The Record therefore will not convert a strong leaderboard position into the sentence:
WeatherNext 3 is the world’s most accurate weather model.
The available evidence is more specific than that.
AI has not replaced meteorology
The arrival of learned weather models has created an easy but misleading story: artificial intelligence replacing physics.
WeatherNext 3 does not justify that story.
The system exists inside a much larger forecasting ecosystem built from physical observations, conventional numerical weather prediction, data-assimilation systems, meteorological expertise, and official warning institutions.
Brightband explicitly labels the machine-learning systems it evaluates as experimental and states that they do not replace official forecasts, alerts, or warnings.
The important historical development is not disappearance.
It is competition and hybridization.
Learned systems are becoming fast and capable enough to sit beside conventional forecasting pipelines, feed products, and potentially influence operational decisions.
That may eventually change which parts of forecasting are performed by numerical solvers, learned models, or combinations of both.
WeatherNext 3 is evidence that the boundary is moving.
From research model to infrastructure
The other reason this release matters is distribution.
Google says WeatherNext 3 is being integrated into Search, Gemini, Maps, Google Maps Platform, Cloud, and Earth Engine.
That places the model on a path from research paper to infrastructure.
A weather model embedded in a consumer product can shape choices about travel, outdoor activity, transportation, logistics, energy use, agriculture, and emergency preparation even when the person making the decision never knows the model’s name.
The consequences of a forecasting system therefore depend on more than benchmark accuracy.
They depend on where the model is deployed, how uncertainty is communicated, how quickly updates propagate, and how users respond to its predictions.
Integration into consumer and developer products can shorten the path between model output and human action.
Why this enters the Record
Artificial-intelligence history cannot be reduced to conversations with language models.
Some of the most consequential systems will predict parts of the physical world.
Weather is an especially revealing case because prediction quality has direct consequences for public safety, agriculture, transportation, energy, logistics, insurance, and daily life.
WeatherNext 3 enters the Record because three changes arrived together:
fresh satellite observations are being incorporated directly into the learned forecasting system;
new forecasts can be initialized every hour;
and Google is moving the resulting system into widely used products and data services.
Independent live benchmarking gives us an external signal that the system is competitive, while the benchmark’s own limitations prevent a universal accuracy claim.
That combination — technical change, independent measurement, and deployment — is more historically useful than another company superlative.
What changed today
A major AI weather model began operating on a faster observational rhythm.
It can take in live satellite observations, initialize a new forecast every hour, and produce selected surface fields at substantially finer resolution than its predecessor.
It is also moving into widely used consumer products and developer services.
The model remains imperfect.
Its strongest comparative claims depend on specific benchmarks and variables.
Official meteorological agencies and numerical weather prediction remain essential.
But the role of learned models in weather forecasting is becoming less hypothetical.
They are moving into the machinery through which people encounter the forecast.