Google DeepMind and Google Research launched WeatherNext 3 on Thursday, describing it as the most advanced and accurate global weather AI model the company has ever built. The new system produces a fresh forecast every hour instead of every six, and resolves key surface variables such as temperature and moisture at a 5-kilometer grid — a picture Google says is roughly five times sharper than its predecessor, WeatherNext 2.
Unlike most model launches, this one is already woven into products used by billions. According to Google, WeatherNext 3 now powers weather results across Search, Gemini, Maps, Google Maps Platform, and Google Cloud. For followers of the latest AI news and research, weather forecasting has quietly become one of machine learning's most commercially successful frontiers — a domain where AI systems are no longer research demos but the default engine behind everyday answers.
A fivefold jump in resolution
A forecast's usefulness often comes down to detail: how finely a model resolves both space and time. WeatherNext 3 generates hourly forecasts at multiple spatial resolutions. Surface temperature and moisture are rendered at 5 kilometers, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers.
WeatherNext 2, by comparison, produced forecasts on a 25-kilometer grid in 6-hour increments. In a comparison published alongside the announcement, the new model's 2-meter temperature forecast over the United Kingdom resolved intricate local topography, avoiding the pixelated, over-smoothed thermal representations that the older grid produced.
Learning from live satellites, not just simulations
The deeper change is what the model learns from. Most AI weather models, including WeatherNext 2, are trained on outputs from numerical weather prediction (NWP) — the supercomputer-driven physics simulations that have anchored operational meteorology for decades. Those simulations carry a six-hour data lag, which introduces biases into fast-changing variables such as rain and surface temperature.
WeatherNext 3 takes a different route. It ingests a live mosaic of global geostationary satellite imagery refreshed every hour, alongside traditional historical analysis, and feeds both into a single Functional Generative Network (FGN) — a mesh transformer architecture that outputs dense gridded fields, discrete cyclone tracks, and station-level predictions natively. Because every hourly forecast is grounded in the most recent satellite observations, Google says the model can track fast-changing weather with far greater precision, with some of the largest gains in precipitation — the rain-and-snow predictions people actually plan their days around.
From grid cells to cyclone tracks
The FGN's outputs are notably diverse for a single model. Beyond gridded fields, it emits discrete cyclone tracks and sparse station-level coordinates as first-class predictions rather than post-processed add-ons. That matters for forecasters monitoring hurricanes and typhoons, where track errors translate directly into evacuation decisions. Google's previous WeatherNext generation drew attention in August for a breakthrough in tropical cyclone forecasting, and the new architecture extends that line of work.
Where the forecasts will appear
Google is rolling WeatherNext 3 out across Search, Gemini, Maps, Google Maps Platform, and Cloud. The company also highlighted two additions aimed at professional users: sharper precipitation forecasting and new clean energy variables, outputs designed for the renewable-energy sector where wind and solar operators plan around hourly atmospheric shifts.
Enterprises can integrate the forecasts through Google Cloud, which positions WeatherNext 3 not just as a consumer feature but as a data product — one that agriculture firms, logistics companies, and grid operators can build on directly.
How the accuracy claim was measured
Google's claim that WeatherNext 3 is the most accurate global weather model to date rests not only on internal testing but on independent live evaluations conducted by Brightband, a weather AI company that runs continuous benchmarks of forecasting systems. Inviting third-party evaluation is becoming a pattern for Google's science models, and it addresses a longstanding criticism that AI weather papers report results on curated historical sets rather than live conditions.
Why this matters beyond the umbrella app
Google's announcement frames weather as an input to decisions at every scale — from grabbing an umbrella to managing heatwaves and droughts that ripple through agriculture, global supply chains, clean energy production, and national economies. In recent years, AI forecasters have earned their place in that pipeline by generating predictions faster than traditional methods while matching or beating their accuracy, according to Google.
The competitive landscape is heating up as well. Meteorological agencies and technology companies alike have shipped AI forecasting systems in recent years, and hourly, satellite-grounded forecasts raise the bar for all of them. The question for the next year is whether national weather services begin treating models like WeatherNext 3 as primary sources rather than experimental supplements — and whether Google's hourly cadence becomes the industry's new baseline.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
