DeepMind WeatherNext 3 Hits Hourly 5km Forecasts

Google DeepMind and Google Research on 3 September 2026 introduced WeatherNext 3, a global weather AI that ingests live geostationary satellite mosaics and issues a fresh forecast every hour. Independent live evaluations by Brightband called it the most advanced and accurate global weather AI to date, according to Google's announcement. Temperature and moisture fields run at roughly 5 kilometers, other surface variables at 10 kilometers, and atmospheric wind fields at 25 kilometers — about five times sharper than WeatherNext 2's 25-kilometer fields on a six-hour update cycle.

Resolution gains and training data mix

WeatherNext 3 trains on sparse station observations, NASA IMERG precipitation products, and Google's own precipitation reanalysis, which helps the model stay grounded where satellite-only methods historically drifted. Google highlighted precipitation continuous ranked probability score improvements of up to 60 percent versus IMERG, 30 percent versus MRMS, and 10 percent versus rain gauges at early lead times. For planners looking a day or more ahead, the company cited up to 50 percent more accurate precipitation — with the largest gains claimed in historically underserved regions across Latin America, Africa, and Asia-Pacific.

Clean-energy variables are part of the same release: 100-meter wind fields sized for turbine hub heights, cloud cover, and solar radiation. That packaging is aimed at grid operators and project developers who already buy weather APIs rather than at casual phone users alone. Readers following Google's research-to-product pipeline may also compare the cadence with other DeepMind-adjacent shipping stories covered in Google Gemini 3.8 Flash Fairwind Cyber.

Where the forecasts will show up

Google said WeatherNext 3 is rolling into Search, the Gemini app, Maps, the Maps Platform Weather API, Earth Engine, BigQuery, and Google Cloud Storage. That distribution list matters more than a single demo video: enterprise buyers can query the same underlying fields from analytics warehouses while consumers see quieter improvements inside familiar Google surfaces. Hourly refresh from satellite mosaics also changes operational expectations for nowcasting-style use cases that previously waited on slower global model cycles.

Even so, Google's disclaimer is unambiguous: official warnings still come from national meteorological services. Media, energy desks, and municipal emergency managers should treat WeatherNext 3 as decision support layered on top of those authorities, not as a replacement alert channel. That framing mirrors how other AI weather systems have been introduced — capability first, liability boundaries second.

Why hourly five-kilometer fields matter

A five-kilometer temperature and moisture grid updated hourly is sharp enough to matter for flash-flood risk pockets, solar ramp forecasts, and urban heat planning that coarser six-hour fields often smeared together. The 25-kilometer wind fields remain coarser aloft, so aviation and severe-storm specialists will still blend WeatherNext outputs with dedicated mesoscale and ensemble systems. Brightband's live-eval praise is a strong early signal, but operational agencies will want multi-season verification before rewriting playbooks.

For product teams, the immediate work is integration testing: confirm latency on Maps Platform Weather API calls, validate BigQuery schema stability, and measure whether Gemini app summaries cite the new fields without overconfident language. Energy traders should backtest 100-meter wind and solar radiation series against local SCADA before sizing positions. Humanitarian and development organizations focused on LatAm, Africa, and APAC should ask Google for region-specific skill scores rather than relying only on global averages.

File this as research news dated 3 September 2026 with product distribution already underway across Google's consumer and cloud surfaces. WeatherNext 3's headline is not a single benchmark number but the combination of hourly cadence, finer surface resolution, and claimed precipitation skill in underserved regions — all under the standing rule that national met services own official warnings. Related policy and industry context on how labs ship high-stakes AI appears in EU AI Office First RFIs to Frontier Labs.

Analysts writing follow-ups should separate three claims carefully: Brightband's independent live evaluation language, Google's CRPS precipitation percentages, and the product rollout list. Mixing those threads into one undifferentiated hype paragraph is how weather AI coverage loses credibility with working forecasters. PromptCrates will watch for peer-reviewed verification papers and third-party agency comparisons as WeatherNext 3 spends its first full seasons in production.

Until those longer verification windows close, the responsible editorial stance is measured: WeatherNext 3 looks like a genuine resolution and cadence step over WeatherNext 2, Brightband's early live evals are encouraging, and Google's own warning disclaimer remains the correct safety rail for any public-facing rewrite of this story.

Municipal planners and insurance modelers should also ask how often WeatherNext 3 fields are versioned in BigQuery and Earth Engine, because silent schema changes can break automated dashboards that assume yesterday's column names still mean the same physical quantity tomorrow.

Sources

• Introducing WeatherNext 3 — Google DeepMind / Google Research blog, 3 September 2026