Google launches AI-powered model for global weather forecasting

Google has launched WeatherNext 2, an AI-powered model for global weather forecasting developed by its DeepMind and Research teams.

The model is available on Earth Engine and BigQuery, with early access on Google Cloud’s Vertex AI.

WeatherNext 2 is already used in Search, Gemini, Pixel Weather, and Google Maps Platform’s Weather API, with Maps support coming soon.

Google said WeatherNext 2 can generate forecasts up to eight times faster and at a higher resolution than earlier versions.

The model can produce hundreds of possible weather scenarios from a single input in under a minute on a single TPU.

It also outperforms the previous model in almost all tested variables and lead times, using a new Functional Generative Network approach.

🔗 Source: Google

🧠 Food for thought

Implications, context, and why it matters.

Google says WeatherNext 2 beats its prior model across variables and lead times, yet it has not shared third-party checks against the European Centre for Medium-Range Weather Forecasts (ECMWF) High Resolution (HRES) numerical model (a physics-based forecast) or AI systems like GraphCast and Pangu-Weather (AI weather forecasting systems).
Recent work finds ECMWF’s HRES beats state-of-the-art AI models including GraphCast and Pangu-Weather on record-breaking extremes 1. A 2024 study tested five global AI weather models and saw performance vary by region and phenomenon. FengWu (an AI global weather model) ranked best overall, while multi-model ensembles (combined forecasts from multiple models) matched top single-model runs 2.
Today’s AI systems train on numerical model outputs (simulations from physics-based numerical weather prediction) and reanalysis data (historical observations blended with models to create a consistent weather record). They have not replaced physics-based forecasting. Performance can slip on unprecedented extremes outside training bounds 3.

WeatherNext 2 now powers Google Maps Platform’s Weather API. Teams building location services, logistics tools, or outdoor recreation apps can add forecasts without running infrastructure.
Google Maps Platform uses pay-as-you-go pricing with free monthly caps (up to 10,000 free calls per product monthly across subscription tiers). Volume discounts start at 20% above 100K monthly requests 4. Pricing sits under billing. Docs cover quotas, adjustments, and terms 5.
Review these to plan unit economics (per-request costs and margins) and scalability. Rivals such as OpenWeather (a third-party weather data provider) publish tiered plans that start with 1,000 free daily API calls and scale to billions of monthly calls with defined rate limits (maximum requests allowed within a time window) 6.