The useful change is the refresh cycle

A weather forecast is not only a prediction. It is a snapshot of what a system knew when it started. Google says WeatherNext 3 narrows that gap by ingesting low-latency geostationary satellite mosaics and producing a new global forecast every hour. Earlier operational WeatherNext work refreshed on a six-hour cycle.

That sounds like an engineering detail, but it is the heart of the release. Rain bands, coastal winds and fast-growing storms do not wait politely for a model update. A forecast that can take in newer observations more often may be useful even before it is more accurate on a benchmark.

Google DeepMind and Google Research announced WeatherNext 3 on 3 September. A paper posted to arXiv that day describes the same broad approach: a probabilistic global system that takes raw observations, including satellite data, rather than relying only on the analysed weather fields that many AI forecast models use. It is a preprint, not a peer-reviewed result.

More detail, but not one magic resolution

The company describes several resolutions rather than one headline number. WeatherNext 3 forecasts selected single-level surface variables, including temperature and moisture, at 5 kilometres; other surface variables at 10 kilometres; and atmospheric variables such as wind speed at 25 kilometres. Its developer documentation describes a 15-day global probabilistic forecast with 64 ensemble members, initialized hourly.

That distinction matters. Forecast skill is not a single property that rises evenly as a grid gets finer. Some parts of the atmosphere can be usefully represented at one scale, while local rainfall and surface conditions need different observations and evaluation methods. The model also produces station-level predictions, which Google says are conditioned on local geography.

Google reports that the new system is roughly five times sharper than WeatherNext 2 for the surface fields it compares. In its announcement, it also reports improvements for precipitation against several reference datasets. Those numbers are the company’s results. The open paper gives researchers a method to inspect, but it does not make a newly released system independently settled science overnight.

Where it is becoming operational

WeatherNext 3 is not being presented only as a research demo. Google says it is beginning to power weather experiences in Search, the Gemini app, Google Maps, the Google Maps Platform Weather API and Earth Engine. It also says developers and researchers can access forecast datasets through BigQuery, Earth Engine and Cloud Storage, and can run tailored inference through Google Cloud.

The model’s energy-related outputs are part of the practical case. Google says it forecasts 100-metre wind speed, cloud cover and solar radiation, variables that can help estimate wind and solar generation. That could be useful for grids and renewable operators, but the release does not demonstrate a particular operational saving or a public-sector deployment outcome.

For everyday users, the immediate effect will be difficult to separate from the rest of a weather product. A forecast may mix model output, alerts, local observations and interface choices. Google says users planning a day or more ahead will see more accurate precipitation forecasts. That is a product claim to test over changing weather, not a promise that every local forecast will be right.

Why the observation layer matters

Traditional numerical weather prediction is built around physics simulations and data assimilation, running on large computing systems. AI forecasters can be much faster, but many have started from the same analysed fields and update rhythms. WeatherNext 3 tries to bring satellite observations and station measurements closer to the model itself.

The trade-off is not that physics disappears. The preprint compares a learned model with existing forecast systems and evaluation datasets; it does not argue that one method makes the other obsolete. The sensible reading is that operational forecasting is becoming a hybrid environment: observation networks, physical models, learned systems and forecasters all still matter.

That is also why a national meteorological agency remains the right place to look for a severe-weather warning. Google includes that reminder in its own release. A high-resolution AI forecast may help people prepare, but it does not carry the legal authority, local responsibility or emergency communication role of an official warning service.

What is confirmed, what Google says, and what is open

Confirmed: Google announced WeatherNext 3 on 3 September 2026, published a supporting arXiv preprint, and made developer documentation available. The public materials describe hourly initialization, satellite inputs, multiple forecast resolutions and planned or current integration across several Google products and Cloud services.

Google’s claims: WeatherNext 3 is its most advanced and accurate global weather model; its higher-frequency, higher-resolution approach improves forecast performance; and users will receive more accurate precipitation forecasts at longer lead times. Google also describes the system as useful for renewable-energy planning and areas underserved by high-resolution forecasting.

Open questions: how results hold across seasons, regions and rare high-impact events; which evaluation results outside Google will emerge as adoption grows; how downstream products communicate uncertainty; and how forecasters and public agencies will use these AI outputs alongside established numerical models and warning systems.

Sources

  1. Google — Introducing WeatherNext 3, our most advanced and accurate global weather AI modelPrimary Google announcement, published 3 September 2026. Source for the launch, stated product integrations, technical descriptions and reported performance claims.
  2. WeatherNext 3 preprint — Increasing resolution and performance of global weather models with raw observationsTechnical preprint submitted 3 September 2026. It describes the model inputs, hourly operation and reported evaluation approach; it has not yet been peer reviewed.
  3. Google for Developers — WeatherNextPrimary developer documentation for the stated 15-day probabilistic forecast, ensemble configuration, data access and inference options.