One Model Now Sits Behind Three Everyday Surfaces
Google DeepMind introduced WeatherNext 3 on September 3, 2026, and shipped it directly into Google Search, the Gemini app, Google Maps, the Maps Platform Weather API, Earth Engine, and BigQuery. The model generates a fresh global forecast every hour instead of every six, ingesting live satellite data as it arrives rather than waiting for the next scheduled run. Surface temperature and moisture resolve to a 5 kilometer grid, other surface variables to 10 kilometers, and atmospheric variables such as wind speed to 25 kilometers, a jump Google describes as roughly five times sharper than its predecessor, WeatherNext 2.
Google's own post states the model delivers up to 50 percent more accurate precipitation forecasts for planning a day or more ahead, with the largest gains in regions where forecasting has historically been weakest. It also names two specific industrial uses: turbine output estimation from 100 meter wind-speed predictions for renewable energy operators, and cloud-cover and solar-radiation forecasting for solar farms.
One Number Is Independently Checked, and It Is Not the One in the Pitch
WeatherNext 3 is not purely self-graded. Brightband, an independent weather-forecasting company, runs Operational WeatherBench, a live leaderboard that continuously scores operational AI and physics-based weather models against each other under real-world conditions. According to Brightband's own tracking, WeatherNext 3 posted the lowest 2 meter temperature error on 26 of the last 30 days of August 2026, ahead of both ECMWF's IFS and NOAA's GFS, the two reference physics models meteorologists have relied on for decades. That is a genuine third-party result, run by a party with no stake in Google's success, and it deserves to be treated as real evidence rather than dismissed alongside the rest.
The gap sits one layer down. Brightband's leaderboard tracks temperature skill. The numbers Google actually leads its own announcement with, and the ones a renewable-energy or agricultural operator would need before trusting the model for a real decision, are precipitation-accuracy comparisons: up to 60 percent better against IMERG satellite data, 30 percent against MRMS radar data, and 10 percent against ground rain gauges. Those three figures do not come from Brightband. They come from Google DeepMind evaluating its own model against datasets it selected.
Verified Temperature Skill, Self-Reported Precipitation Skill
| Claim | Who measured it |
|---|---|
| Lowest 2 meter temperature error, 26 of last 30 days (August 2026) | Brightband, independent third party |
| Up to 60 percent better than IMERG on precipitation | Google DeepMind, self-reported |
| Up to 30 percent better than MRMS on precipitation | Google DeepMind, self-reported |
That split matters most for the two industrial uses Google names in its own post: turbine output estimation from 100 meter wind-speed predictions, and cloud-cover and solar-radiation forecasting for solar farms. Neither wind output nor solar irradiance is temperature. The metric an operator would actually need independently checked, precipitation and cloud-cover accuracy, sits on the self-reported side of the table, while the metric that happens to be independently verified, temperature, is not the one those two use cases depend on.
None of this means the precipitation numbers are wrong. DeepMind has a strong track record on weather, and Brightband's own ranking is evidence the underlying model is genuinely capable. It means a wind-farm or solar-farm operator dispatching output on this data is relying on a mix of one audited score and one vendor-reported score, presented in the same announcement with no indication of which is which.
Read next: The First AI Benchmark Neither Side Could See | Gemini 3.7 Flash Halves Its Price Until January
