Google Launches WeatherNext 3, Claims 50% More Accurate Precipitation Forecasts

Google today released WeatherNext 3, a new AI weather model that the company calls its "most advanced and accurate global weather AI model" to date. Built by Google DeepMind and Google Research, the model is already rolling out to Search, Maps, and the Gemini app, with Google claiming up to 50% more accurate precipitation forecasts for day-ahead and longer-range planning.

What changed

Previous weather AI models, including Google's own WeatherNext 2 from November 2025, were trained on output from numerical weather prediction (NWP) systems — physics-based supercomputer simulations that carry a six-hour data lag. That lag introduces bias for fast-moving variables like rainfall and surface temperature.

WeatherNext 3 instead trains on a "mosaic of live, global geostationary satellite data," according to Google, allowing the model to learn directly from real-time atmospheric observations rather than delayed simulation output. Google says this produces a "continuously updating view of the atmosphere" and enables hourly forecast generation.

Resolution jumps significantly over the prior version:

• Temperature and moisture: 5-kilometer resolution
• Surface variables: 10-kilometer resolution
• Wind speed: 25-kilometer resolution

WeatherNext 2 topped out at 25-kilometer resolution with 6-hour update increments. Google describes the new model's output as "roughly five times sharper."

Precipitation accuracy claims

Google trained WeatherNext 3's precipitation forecasting specifically on NASA's IMERG (Integrated Multi-satellite Retrievals for GPM) dataset and an internal global precipitation reanalysis built from satellite radar. According to Google, evaluations against baseline datasets show Continuous Ranked Probability Score (CRPS) improvements of:

• Up to 60% against IMERG
• 30% against MRMS (Multi-Radar Multi-Sensor)
• 10% against rain gauge measurements at early lead times

For consumer-facing day-ahead and multi-day forecasts, Google claims up to 50% more accurate precipitation predictions, with the largest gains concentrated in regions — Latin America, Africa, and Asia-Pacific — that Google says have historically lacked high-resolution regional forecasting due to the supercomputing cost of traditional NWP models.

None of these accuracy figures have been independently verified; they come directly from Google's own benchmarking against the cited baseline datasets.

Beyond consumer forecasts

Google says WeatherNext 3 also generates forecasts aimed at renewable energy operators, including 100-meter wind speed predictions (roughly turbine height) for wind-energy output estimates, along with high-resolution cloud cover and solar radiation forecasts intended to help solar farms estimate ground-level light exposure.

What this means

The shift away from NWP-derived training data toward live satellite observation is the more consequential technical change here — it addresses a structural bottleneck (the six-hour simulation lag) rather than just adding compute or parameters. If Google's CRPS figures hold up under independent scrutiny, a 60% improvement against IMERG for medium-range precipitation would be a meaningful jump for an area where AI weather models have historically struggled relative to physics-based systems for short-fuse events like convective storms.

The geographic framing is also notable: regions historically underserved by regional NWP models due to compute cost stand to gain the most, since WeatherNext 3's satellite-based approach doesn't require the same localized supercomputing infrastructure. That could matter more in practice than the marginal gains for already well-forecast regions like North America and Europe. As with all self-reported model benchmarks, the real test will be how these numbers hold up against independent verification once meteorologists and third parties can evaluate live forecast performance rather than Google's internal backtests.