---
type: analysis
slug: research-weathernext-3-s-50-better-rain-forecasts-claim-comes-from-google-alone-but-the
title: "WeatherNext 3's \"50% better rain forecasts\" claim comes from Google alone, but the shift to training on observations is real"
description: "Google DeepMind's new weather model refreshes every hour from satellite data and learns partly from raw observations. Its headline precipitation gains come from Google's own scoring and depend on which \"ground truth\" you score against: up to 60% against a dataset the model was trained on, but only about 10% against rain gauges."
published_at: 2026-10-23T23:00:00.000Z
topics: ["google","research"]
source_urls: ["https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3","https://arxiv.org/html/2609.03582v1","https://www.brightband.com/company/news/weathernext-3-on-operational-weatherbench/","https://developers.google.com/weathernext","https://arxiv.org/abs/2609.03582"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
On 3 September 2026, Google DeepMind and Google Research released WeatherNext 3. They describe it as their most accurate global AI weather model, and it now powers weather in Search, Gemini, Maps, the Maps Platform Weather API and Earth Engine. The marketing focuses on precipitation, an area where AI models have tended to blur rainfall or miss storm edges, with Google claiming "up to 50%" more accurate rain forecasts a day or more ahead. Other headline claims are an hourly refresh instead of six-hourly and 5 km surface resolution, about five times sharper than WeatherNext 2.

This matters to builders for two reasons. First, the model is being sold as infrastructure: forecast data is available in BigQuery, Earth Engine and Cloud Storage, and Google offers custom inference on its accelerators. Second, it marks a change in method. Instead of learning only to imitate physics-based reanalysis, the model learns from raw observations as well. Taking the precipitation number at face value without knowing what it was scored against would be a mistake for anyone building flood, logistics or energy products on top of it.

## What we checked

- **Does this cluster point to a broader architecture transition that a single release note can't capture?** partly confirmed. Yes in part. WeatherNext 3 moves from models trained only on NWP or analysis output, which runs about 6 hours behind, to a model that also ingests hourly geostationary satellite imagery and learns to predict station, satellite-precipitation and cyclone observations directly. It still uses analysis data as its foundation, so this is a hybrid step rather than a full move to observation-only models. [1][2]
- **Has anyone independently verified the 'up to 50% reduction in precipitation error' claim?** refuted. No. The 50% figure comes from Google's blog, and I couldn't find it in the paper text I read. The paper reports CRPS gains of up to 60%, 30% and 10% at early lead times, depending on which verification dataset is used. Brightband's independent leaderboard post covers 2m temperature and doesn't mention precipitation. [1][2][3]
- **Does WeatherNext 3 lead independent live evaluation on core surface variables?** partly confirmed. Brightband's Operational WeatherBench ranked the WeatherNext 3 ensemble as the leader in August 2026. It had the lowest global 2m-temperature RMSE (days 1–7) on 26 of 30 days, but it only narrowly beat WeatherNext 2. That covers one month and mainly one variable. [3]

## What we found

- Google's blog says WeatherNext 3 ingests live hourly geostationary satellite mosaics plus historical analysis data through a single Functional Generative Network mesh transformer. It outputs gridded fields, cyclone tracks and station-level predictions. Resolution is 5 km for surface temperature and moisture, 10 km for other surface variables and 25 km for atmospheric variables, and a new forecast is issued every hour. [1]
- The paper (arXiv 2609.03582, Rasp et al., 3 September 2026) reports precipitation CRPS reductions against WeatherNext 2 and ECMWF ENS of up to 60% when scored against IMERG, 30% against MRMS radar and 10% against rain gauges, all at early lead times. The paper itself says IMERG is not an independent ground truth for WN3, because IMERG is one of its training targets. The cleaner checks show much smaller gains: MRMS covers only a limited region, and gauges are unevenly spread. [2]
- The paper lists real limitations. It evaluates precipitation only up to moderate amounts (4 mm/6h) and leaves extreme rainfall to future work. The real-time sample is only six weeks and its precipitation results are noisy. Individual ensemble members show hexagonal mesh artifacts, the station head overfits, and cyclone intensity spread is under-dispersed. This sits awkwardly with the marketing about catching severe storm boundaries. [2]
- The paper reports 64-member ensembles out to 15 days. Each ensemble member takes about 6.3 minutes on 4 TPUv5p chips. Training both models took about 4.8 TPUv4 chip-years and 8.9 TPU7x chip-years. Starting forecasts every hour instead of every six hours gives roughly 2–3 hours of effective lead-time gain. [2]
- Brightband, an outside forecasting company, runs the Operational WeatherBench live leaderboard. It named WeatherNext 3 the leader for August 2026, with the lowest global 2m-temperature RMSE on 26 of 30 days. It said the model only narrowly beat WeatherNext 2 and AIFS-ENS. The post does not evaluate precipitation. [3]
- Google's developer page offers WeatherNext forecasts through Cloud Storage (Zarr), Earth Engine and BigQuery, plus custom on-demand inference on Google Cloud. It doesn't say whether WeatherNext 3 weights are open, and gives no pricing or license terms. The paper also doesn't say whether weights will be released. [4][2]

## What it means

The technical shift holds up. WeatherNext 3 shows that a global AI model can take in low-latency satellite data and learn station and precipitation targets directly, without waiting about six hours for a reanalysis. You get hourly forecasts and point predictions even at stations it never saw in training. If your product depends on freshness, such as short-term solar, wind or nowcast-style rain alerts, the hourly cycle and 10 km hourly precipitation are worth testing. Treat this as Google-hosted data, not a model you can run yourself: access goes through Google Cloud, and there are no public weights or pricing yet.

Read the precipitation headline as a range that depends on how you score it, not as a single number. The largest gain (60%) is measured against a dataset the model was trained to reproduce. Against radar it is about 30%, and against rain gauges about 10%, and only at early lead times and moderate rainfall. Extreme rainfall isn't evaluated, and the only outside live evaluation so far covers temperature. Before relying on it for flood or severe-weather decisions, check it against your own gauge or radar data for your region and the thresholds you care about. Keep official warnings from national weather services in the loop, as Google itself advises.

## Still open

- Where exactly does Google's 'up to 50% more accurate a day or more ahead' precipitation figure come from? Which metric, verification dataset, baseline and lead time? I couldn't find it in the paper text I read.
- How does WeatherNext 3 perform on heavy and extreme precipitation (above 4 mm/6h) and on severe convective storms?
- Will Brightband or another independent group evaluate precipitation on a live leaderboard?
- What will the pricing, license terms and any open-weight release for WeatherNext 3 be?

## Sources

1. [Introducing WeatherNext 3 (Google blog)](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3) (primary)
2. [WeatherNext 3 paper, full HTML v1](https://arxiv.org/html/2609.03582v1) (primary)
3. [Brightband: WeatherNext 3 on Operational WeatherBench](https://www.brightband.com/company/news/weathernext-3-on-operational-weatherbench/)
4. [WeatherNext developer page](https://developers.google.com/weathernext) (primary)
5. [WeatherNext 3: Increasing resolution and performance of global weather models with raw observations (arXiv abstract)](https://arxiv.org/abs/2609.03582) (primary)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- topic hub: [Research](https://newruntime.com/topics/research.md): Explore the research topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)