WeatherNext 3: Inside Google DeepMind's Hourly, 5-Kilometer AI Weather Model

5 kilometers and every hour: that's the sharpest resolution and the new cadence of WeatherNext 3, the AI weather model Google DeepMind launched September 3, 2026 — roughly five times sharper overall than its predecessor, per Google, paired with a claimed 50% jump in precipitation accuracy.

Some searchers type the product name as two words — "weather next 3" — without realizing WeatherNext 3 is a single branded model name, not a forecast for three days out. Here's what Google actually shipped, what the accuracy claim does and doesn't mean, and why the model's new wind and solar forecasts matter more commercially than the consumer weather-app upgrade.

WeatherNext 3 explained: what Google's "weather next 3" model actually does

WeatherNext 3 is a global AI weather model from Google DeepMind and Google Research that generates forecasts every hour at multiple resolutions: 5 kilometers for key surface variables like temperature and moisture, 10 kilometers for other surface variables, and 25 kilometers for atmospheric variables such as wind speed. Google says that adds up to a global picture roughly five times sharper than WeatherNext 2, which produced forecasts on a 25-kilometer grid in 6-hour increments. Google calls it the company's most advanced and accurate weather model yet, and it began feeding live weather results in Google Search, Gemini, Google Maps, and the Maps Platform Weather API as soon as it launched.

The technical shift behind the accuracy claim is training data, not just added compute: WeatherNext 3 ingests live geostationary satellite mosaics alongside traditional historical analysis, and trains its 5-kilometer forecasts directly on weather-station observations, according to Google's own announcement. Google says most AI weather models, including WeatherNext 2, train on numerical weather prediction output that carries a six-hour data lag; the live satellite feed is what allows a fresh forecast every hour. TechCrunch notes that WeatherNext 3 still relies on national weather datasets, so this is not yet full direct data assimilation.

WeatherNext 3 vs WeatherNext 2: what actually changed

The headline upgrade is resolution and speed, but WeatherNext 3 also adds forecast variables its predecessor didn't cover. The table below compiles the figures Google, TechCrunch, and 9to5Google reported on the model at its September 3, 2026 launch.

Compiled from Google's announcement, TechCrunch, and 9to5Google reporting, September 2026.

What else Google disclosed — and what's disputed

Behind the "up to 50%" consumer figure, Google's announcement reports precipitation gains measured as Continuous Ranked Probability Score (CRPS) improvements of up to 60% against NASA's IMERG satellite data, 30% against MRMS radar data, and 10% against rain gauges at early lead times. The model also outputs discrete cyclone tracks and forecasts for specific weather-station locations, and TechCrunch reports it has 2.4 times as many parameters as WeatherNext 2. Researchers, developers, and businesses can query the hourly forecast data in BigQuery and Earth Engine or bulk-download it from Google Cloud Storage.

One claim is contested. Google calls WeatherNext 3 the "first" AI model to directly incorporate raw observations for a high-resolution global forecast, but AI weather startup WindBorne says, per TechCrunch, that its WeatherMesh 6 model has been incorporating raw observations from its weather balloons and other sources since late 2025; asked about that, Google pointed out that its forecasts are higher resolution across the globe. Google also notes that for official forecasts, severe weather warnings, and public safety advisories, people should rely on their local meteorological agency or national weather service.

The energy angle: why grid operators matter more than umbrellas

WeatherNext 3's most commercially interesting expansion isn't the consumer accuracy bump — it's the new wind-speed forecast at 100 meters, roughly turbine height, plus cloud cover and solar radiation data. Accurate wind and solar output prediction is a real, persistent pain point for grid operators managing intermittent renewable capacity, and this puts Google in loose competition with specialized energy-forecasting vendors, not just consumer weather apps like AccuWeather or Apple Weather.

Grid capacity already limits how fast new AI data centers can come online, so a model that helps operators time renewable output more precisely has an obvious adjacent use case: matching AI compute demand to periods of high wind or solar generation. Whether Google turns this into a distinct enterprise product, separate from the free consumer integrations in Search and Maps, is the detail worth watching — that's the difference between a research showcase and a new revenue line.

What the 50% accuracy claim misses

Independent third-party validation is on the record. Brightband, a weather-forecasting company unaffiliated with Google, runs Operational WeatherBench (OWB), a live leaderboard that scores leading AI and physics-based global models — including ECMWF's AIFS ensemble and WeatherNext 2 — on the same real-time metrics. In a September 2, 2026 post, Brightband said WeatherNext 3 is the new OWB leader; by Brightband's own count, WeatherNext 3 posted the lowest global 2-meter-temperature error of any tracked model on 26 of the 30 days in August 2026 and led the field more broadly across OWB's headline skill metrics — a genuinely independent result, not a DeepMind-reported number.

That's a meaningfully stronger footing than the single self-reported figure this piece originally flagged, though it isn't a one-to-one confirmation of Google's specific "50% more accurate precipitation forecasts a day or more in advance" claim — OWB scores each model against its own analysis by default (the ECMWF IFS analysis for WeatherNext 3) and always verifies precipitation against ERA5; IMERG, MRMS, and rain gauges are the baselines Google used in its own precipitation evaluation, not OWB's ground truth. DeepMind has a credible track record on weather modeling, including the earlier GraphCast and WeatherNext 2 releases, and WeatherNext 3's lead on a shared, independently run leaderboard is the strongest outside evidence yet that the architectural shift translates into real accuracy gains — not just a marketing number.