Google's New Weather AI Tops a Benchmark Run by Ex-Google Staff

TL;DR

Google DeepMind launched WeatherNext 3 on September 3, 2026, a new AI model for forecasting weather worldwide. It maps temperature and moisture at five kilometers of resolution instead of the previous model’s 25 kilometers, and updates hourly using live satellite data instead of every six hours. On rain forecasts, Google says the model is up to 50% more accurate a day or more ahead, and TechCrunch reports it topped a third-party benchmark, beating rival models and national weather services. That benchmark, though, is run by a company two of whose founders previously worked at Google, and a competing weather startup disputes part of Google’s pitch.

What happened

Google DeepMind and Google Research announced WeatherNext 3 on September 3, 2026, calling it their most advanced and accurate global weather AI model to date. The model produces key surface variables such as temperature and moisture at a native 5-kilometer resolution, against the 25-kilometer grid of its predecessor WeatherNext 2 — a change Google describes directly: “Overall, this provides a global weather picture roughly five times sharper than our previous model, WeatherNext 2, which produced forecasts on a 25-kilometer grid in 6-hour increments.” The new model also ingests real-time geostationary satellite data and updates its forecasts hourly, rather than every six hours.

On precipitation specifically, Google reports a Continuous Ranked Probability Score improvement of up to 60% against the IMERG satellite dataset, 30% against the MRMS radar dataset, and 10% against rain gauges for early lead times — a statistical measure of how closely a forecast’s probabilities match what actually happened. For forecasts a day or more out, the company says people will see up to 50% more accurate precipitation forecasts, with the biggest gains in regions where forecasting has historically been weaker, though it names no specific regions. Google says the model will feed into Search, the Gemini app, Google Maps, the Maps Platform Weather API and Google Earth Engine, and be made available to researchers via BigQuery and Cloud Storage. Google senior staff engineer Samier Merchant told TechCrunch: “This is going to be the first time that some of the core variables feed and power a lot of the Google products.”

Google’s “most accurate” ranking, per its own post, rests on independent live evaluations run by Brightband. TechCrunch reports that on Brightband’s Operational WeatherBench benchmark, WeatherNext 3 beat deep-learning models from Google, Microsoft, Nvidia and the European weather center ECMWF, as well as traditional forecasts from the U.S. National Weather Service and ECMWF. DeepMind staff research scientist manager Ferran Alet explained the underlying approach to weather modeling: “Weather is chaotic, and so small differences really start to perturb massively…Machine learning targets the problem we are really solving, which is approximate noisy physics from incomplete information and finite compute, and so it learns patterns from a lot of data.”

What this means (and what it does not)

The practical stakes are in distribution: Google is building WeatherNext 3 directly into Search, Maps, Gemini and Earth Engine, so any accuracy gain reaches people who never open a weather app.

What does not follow is that Brightband’s leaderboard is a neutral outside check. Brightband was co-founded by Julian Green, who previously ran AI projects at Google X, and Daniel Rothenberg, a meteorologist formerly at Google Research. Rothenberg is also the outside voice TechCrunch quoted favorably on the model’s design: “The idea, with a lot of AI applications, is to try to run tasks as end-to-end as possible.” Nor does Google’s framing of WeatherNext 3 as first to use live observations hold unchallenged: TechCrunch reports that competitor WindBorne’s WeatherMesh 6 has incorporated raw weather observations since late 2025, and that both models still lean on national weather datasets rather than true direct data assimilation — Google’s reply was that its forecasts have higher global resolution. Separately, Gizmodo’s coverage of the launch restated Google’s claims without independent verification and included no quotes from meteorologists or competitors, so it adds no confirmation of its own.

What we still do not know

No source found breaks down WeatherNext 3’s performance against ensemble physics models like ECMWF’s ENS specifically for extreme events — heavy rain, snow, hurricanes — by region, and Google names no specific regions or figures for its claim of “regions where forecasts have historically been less reliable.” Whether WeatherNext 3’s training data leaves geographic blind spots, for instance in areas with sparse weather stations or satellite coverage, is not addressed by any source found. Google has not said whether it will publish its own rolling accuracy metrics; the only ongoing outside tracking identified is Brightband’s leaderboard, whose live rankings could not be checked directly on this pass. The methodology behind the 60%/30%/10% precipitation-accuracy figures — sample size, evaluation period, forecast lead times — is not disclosed in Google’s post, and no technical paper was found to check it. Google’s own WeatherBench 2 benchmark, which compares more than two dozen physics-based and machine-learning models, had not listed WeatherNext 3 results as of this check. And Google’s claim of better snow prediction appears only as an unquantified line in its promotional summary, with no snow-specific accuracy metric given anywhere.