Google's WeatherNext 3: AI Weather Forecasting Just Got Hourly, Global, and 60% Better

What Launched: An AI That Watches the Sky Directly

On September 3, 2026, Google DeepMind and Google Research introduced WeatherNext 3, their most advanced global weather model to date — and by independent live evaluation from Brightband, the most accurate one available anywhere. Unlike traditional forecasting systems that run physics simulations on room-sized supercomputers, WeatherNext 3 is a neural network that learns directly from real-time observations, ingesting a mosaic of raw satellite imagery to produce a fresh, high-resolution forecast every single hour.

That last part is the revolution. Legacy national weather models typically refresh their global forecasts every 6 to 12 hours, and high-resolution regional models are so computationally expensive that most of the planet can't afford to run them at all. WeatherNext 3 sidesteps the supercomputer entirely: because it learns from live satellite data rather than cycling a full physics simulation, it can deliver localized, hourly-refreshed predictions for everywhere — a first for regions across Latin America, Africa, and Asia that have historically been underserved by high-resolution forecasting.

And it's not a lab demo. Starting today, WeatherNext 3 powers the weather experiences in Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine — meaning billions of people just got a forecast upgrade without doing anything.

Why It Matters: The End of Supercomputer Weather

Weather forecasting has been one of the quietest — and most consequential — AI success stories of the past few years. The traditional approach, numerical weather prediction, divides the atmosphere into a 3D grid and solves fluid dynamics equations step by step. It works, but it demands some of the largest supercomputers on Earth and caps how often and how finely you can forecast.

Machine-learning models flipped the economics. Instead of simulating physics, they learn the atmosphere's patterns from decades of reanalysis data, then extrapolate in seconds what takes a supercomputer hours. WeatherNext 3 pushes this further by learning from raw satellite observations directly — the model essentially watches the planet in real time and predicts what happens next, the same way a language model predicts the next token, except the "tokens" are atmospheric states.

The practical consequence: hourly, high-resolution global forecasting no longer requires national-scale compute. A weather startup, an airline, or an agricultural cooperative can access frontier-grade forecasts through an API instead of a data center.

The Numbers: 60% Better Precipitation, 50% Better Planning

Precipitation is the hardest problem in forecasting — and where WeatherNext 3 posts its most striking results. Google trained the model on IMERG, the global precipitation reanalysis built from satellite radar, and evaluated it against multiple baselines:

For everyday users, Google says medium-range planning forecasts — a day or more ahead — now see up to 50% more accurate precipitation predictions, with the biggest gains exactly where forecasts have historically been least reliable. Deciding whether to plan the hike on Saturday or Sunday just became a materially better-informed decision for a few billion people.

The Sleeper Feature: Renewable Energy Forecasting

The announcement's most underrated detail: WeatherNext 3 introduces purpose-built clean energy variables. The model forecasts 100-meter wind speeds — roughly turbine-height — so wind farms can predict output, alongside high-resolution cloud cover and surface solar radiation so solar operators can estimate generation.

This matters because grid operators' hardest problem isn't generating renewable power; it's knowing when that power will arrive. Wind and solar forecasting errors force grids to keep expensive gas plants spinning as backup. An AI model that predicts turbine-height wind and ground-level solar radiation hourly, globally, is directly monetizable — it turns forecast accuracy into fuel savings and grid stability. Expect energy trading desks and utility analytics teams to be among the heaviest users of the new BigQuery and Earth Engine data feeds.

For Developers and Businesses: How to Get the Data

Google is shipping WeatherNext 3 as infrastructure, not just a consumer feature. Three access paths launched today:

• BigQuery and Earth Engine — query global, hourly-updated forecast data directly, ready to join against your own datasets with no model setup required.
• Bulk download from Google Cloud Storage — for teams building pipelines that need full-history or full-globe data.
• Google Maps Platform Weather API — the commercial weather API for consumer and logistics products, now backed by the new model.

If you're building on top of this, the modern stack is straightforward: pull WeatherNext data into your warehouse, then layer AI analysis on top. Tools like NotebookLM for digesting meteorological documentation, Tableau (whose AI-assisted analytics now handle geospatial flows) for visualization, and research assistants like Elicit or Consensus for surveying the climate literature can compress what used to be a data-science team's quarter into an afternoon.

The Bigger Trend: Science Foundation Models Are Having a Moment

WeatherNext 3 is the latest proof that the foundation-model playbook — massive data, learned representations, scale at inference — is escaping the chatbot box and rewriting computational science. The same pattern is playing out across domains: the same day as this launch, the Institute of Foundation Models in Abu Dhabi released K2 Horizon, a fleet of six fully open models spanning 0.9B to 375B parameters, and frontier labs are racing to apply large models to drug discovery, materials science, and seismology.

For the AI tools ecosystem, this is a new category worth watching: AI for science and operations. General assistants like ChatGPT, Perplexity, and Grok already answer weather-adjacent questions with live data — and now the underlying models they cite are becoming AI systems themselves. The layer below the chatbots is being rebuilt by machine learning.

The bottom line: the forecast on your phone got dramatically better today, and the reason is the same technology powering your coding assistant. When AI learns from raw observations of the world — satellite mosaics, sensor feeds, radar — it doesn't just summarize knowledge; it generates new knowledge. WeatherNext 3 is what that looks like when it touches everyone's morning routine.