Google said on September 3, 2026, that WeatherNext 3 would begin powering Search, Gemini, Maps and the Google Maps Platform Weather API. This is a product rollout: Google’s model guide dates the model’s release to August 2026.

For developers evaluating the forecast datasets, the important distinctions are less visible: when a run becomes available, how far it predicts, and whether a platform supplies individual forecasts or statistical summaries.

Hourly runs have different horizons

The model starts a new forecast run every hour, but only the 00:00, 06:00, 12:00 and 18:00 UTC cycles extend 15 days. Intermediate runs stop at 48 hours. A pipeline that always selects the newest run therefore cannot assume it still has a two-week outlook.

Initialization also differs from delivery. Google’s dissemination schedule targets 07:45 UTC availability in Cloud Storage for the 00:00 run, followed by 08:10 in BigQuery and Earth Engine. These are estimates, not guaranteed deadlines. Hourly initialization should not be read as immediate access to freshly observed conditions.

Resolution is variable-specific. Google documents roughly 5-kilometer output for station-trained temperature and dew point, 10-kilometer gridded surface fields, and 25-kilometer atmospheric pressure-level fields. The latter appear only in the four six-hourly cycles. Calling every output a 5-kilometer forecast would misstate the dataset.

Choose the output, then the platform

Google’s Cloud Storage guide describes a full 64-member ensemble: multiple possible forecast outcomes, stored in Zarr v3 arrays. This is also the route for atmospheric pressure-level data. A separate bucket contains summary statistics. The full-ensemble bucket uses Requester Pays, so data transfer can generate charges for the requesting Cloud project.

By contrast, the BigQuery tables and Earth Engine collections expose precomputed surface means and percentiles. They suit SQL joins or geospatial analysis, but are not substitutes for the individual ensemble members when an analysis needs those outcomes. The choice changes the information available, not just the query language.

Operational-data access also requires an allowlist request. Google’s quick-start guide gives a typical approval time of five to seven business days, so the announcement does not mean every developer has immediate dataset access.

Check existing pipelines before switching

WeatherNext 2 users should review field names and accumulation intervals. Google’s migration notes change 2m_temperature to temperature_2m. The one-hour precipitation fields require corresponding daily-total calculations; six-hour accumulations remain available in the full-ensemble store for the main cycles.

Finally, Google classifies WeatherNext as experimental. Its use disclaimer says outputs are not official forecasts or warnings and must not be the sole information source for protecting life or property. Product integration does not remove that boundary.

Graphic: GeethanTech.