Google DeepMind and Google Research announced WeatherNext 3 on September 3, 2026, describing it as their most advanced global weather model. The system takes live geostationary satellite mosaics directly into forecasting, refreshes predictions every hour and produces selected station-calibrated surface variables at up to 5-kilometre resolution. Google says it is being used across Search, Gemini, Maps, Google Maps Platform and Cloud, while forecast datasets are available through Google Cloud Storage, Earth Engine and BigQuery.[1][3][5]
The current status is a product and data rollout, not a replacement for public meteorological services. Google’s documentation says WeatherNext 3 is an automated, experimental system whose outputs are provided as-is for information and research. Users should rely on national meteorological agencies and local authorities for watches, warnings and safety decisions.[3][4]
Key points
- Google announced WeatherNext 3 on September 3, 2026, and its research paper was submitted the same day.[1][2]
- Live geostationary satellite mosaics allow hourly initialization instead of relying only on older analysis fields.[1][2][3]
- The model combines 0.05-degree station outputs, 0.1-degree gridded surface fields and 0.25-degree atmospheric fields.[3]
- Google lists 64 ensemble members, a 15-day horizon for six-hourly cycles and a 48-hour horizon for interim hourly runs.[3]
- Operational datasets are offered through Cloud Storage, Earth Engine and BigQuery, with access requested through an allowlist.[5]
- Forecasts are experimental and are not official severe-weather warnings.[3][4][5]
What happened
WeatherNext 3 is Google’s latest effort to make AI weather forecasting closer to the timing and granularity of observations. The announcement says earlier AI systems commonly learned from numerical-weather-prediction analyses, which can create a lag for rapidly changing variables. WeatherNext 3 instead ingests low-latency satellite observations and initializes a new forecast every hour.[1]
That cadence may help when fronts, convective bands or storms develop between major model cycles. It does not remove atmospheric uncertainty, but it gives the system more recent information before projecting conditions forward. The accompanying paper describes the approach as moving beyond simply emulating traditional data-assimilation, forecasting and post-processing stages.[2]
Google names Search, Gemini, Maps, Maps Platform’s Weather API and Earth Engine as integration points. For developers and researchers, it also describes operational forecast data in Cloud Storage, BigQuery and Earth Engine.[1][5]
How the model is built
WeatherNext 3 uses a multi-resolution design. Google lists station-trained two-metre temperature and dew-point outputs at 0.05 degrees, or about 5 kilometres; gridded surface fields at 0.1 degrees, or about 10 kilometres; and pressure-level atmospheric fields at 0.25 degrees, or about 25 kilometres.[3] The finest figure therefore applies to particular station-calibrated variables, not every forecast field.
The model guide lists live geostationary satellite mosaics and ECMWF HRES analysis among its inputs. Training sources include ERA5 and HRES forecast data, NASA’s IMERG precipitation product, weather-station observations and satellite mosaics.[3] The research paper says it also predicts satellite-derived precipitation, tropical-cyclone information and station observations rather than treating all weather as one regular grid.[2]
The operational schedule differs by run. Google documents 15-day probabilistic forecasts for the 00, 06, 12 and 18 UTC six-hourly cycles, while interim hourly runs have a 48-hour horizon. It lists 64 ensemble members, allowing users to examine a range of outcomes instead of one deterministic line.[3]
Verified performance and context
Google says WeatherNext 3 is about five times sharper overall than WeatherNext 2, which used a 25-kilometre grid at six-hour increments. It also reports up to 50% more accurate precipitation forecasts for planning a day or more ahead, with the largest gains in regions where forecasts have historically been less reliable.[1] These are Google’s published claims, not a guarantee for every location or event.
The paper presents WeatherNext 3 as a new state of the art for probabilistic medium-range forecasting skill and describes its evaluation methodology.[2] Google’s developer documentation separately reports up to a 50% reduction in Brier score and Continuous Ranked Probability Score against numerical-weather-prediction baselines for precipitation.[3] Those metrics measure specific aspects of probabilistic performance; they do not promise that rain or storm timing will always be correct at street level.
The model adds renewable-energy variables, including 100-metre wind speeds, cloud cover and solar-radiation components. These can inform estimates relevant to wind or solar operations, but a forecast variable is not a site-specific production forecast: terrain, equipment, grid constraints and operational decisions require separate analysis.[1][3]
Why it matters
Finer time and space resolution can make forecasts more useful between major model cycles, particularly for fast-changing precipitation and places whose coastlines, valleys or mountains are smoothed in coarser grids. Google says station training is intended to make surface predictions better match ground measurements, and highlights potential value for emergency response, aviation, agriculture and clean-energy operations.[1][2][3]
The access model extends the system beyond Google applications. Operational datasets can be queried or downloaded through cloud surfaces, although real-time access requires an allowlist request. Google says requests are reviewed on a rolling basis and are typically approved within 5–7 business days.[5] Availability is therefore controlled rather than an unrestricted public feed.
What remains uncertain
The existence of WeatherNext 3 is established by Google’s announcement, documentation and research paper. The open questions concern performance in particular situations and how users interpret added detail. Google acknowledges that reanalysis data have biases, precipitation quality depends on training data, and individual samples can contain visible artifacts, including stronger patterns in some station and precipitation outputs.[4]
Google says ensemble models reduce the smoothing associated with deterministic systems but do not eliminate uncertainty. Its documentation notes artifacts in individual samples and discontinuities across six-hour windows in station outputs, even though ensemble statistics remain useful for many applications.[4] Users should therefore examine probabilities and distributions, not rely on a single precise-looking map.
The experimental status also limits public-safety claims. Google says the forecasts are not official warnings and should not be the sole source for protecting life or property.[3][4][5] Independent evaluations, longer operational records and comparisons across regions are still needed to show how consistently reported gains translate into decisions on the ground.
What happens next
Google’s immediate next step is operational dissemination through its products and cloud data surfaces. Developers can request access, choose Cloud Storage, BigQuery or Earth Engine, and use documented variables and forecast cycles.[5] Google also points users to Weather Lab for live visualization and to its paper and benchmarks for evaluation.[1][3]
The practical test will be sustained use: whether hourly satellite initialization improves rapidly evolving forecasts, whether five-kilometre station outputs remain reliable outside well-observed areas, and how the 64-member ensemble informs applications that need risk estimates. Until then, WeatherNext 3 is a documented forecasting upgrade with meaningful technical advances and explicit limits, not a guarantee of perfect local prediction.[2][3][4]
