Published 4 September 2026
Google publicly introduced WeatherNext 3 on 3 September 2026 and is integrating it across Search, Gemini, Maps, Google Maps Platform and Cloud. The model produces forecasts every hour, with key surface variables at 5 km resolution and other surface variables at 10 km. That is a practical jump from WeatherNext 2, which produced forecasts on a 25 km grid in six-hour increments.
The same day’s developer update put Lyria 3.5 Clip and Pro into Gemini API public preview. For businesses, WeatherNext 3 is the more consequential release because teams can use hourly global forecast data without running the model themselves. Lyria 3.5 is a narrower developer-access change for music generation, with $0.04 clips and $0.08 full-song requests.
What WeatherNext 3 changes for energy and operations teams
WeatherNext 3 uses live geostationary satellite imagery together with observations and historical analysis. Google says the model can update every hour and preserve consistency from global atmospheric patterns down to local terrain. Key surface variables such as temperature and moisture are available at 5 km, other surface variables at 10 km, and atmospheric variables such as winds at 25 km.
For renewable-energy planning, the model adds variables that map directly to production: wind at 100 meters, cloud cover and solar radiation. Grid operators can estimate wind and solar output against demand, while logistics and agriculture teams can join weather data to operating locations.
Where can you get WeatherNext 3 data?
- In products: Google Search, the Gemini app, Google Maps and Google Earth Engine, from 3 September 2026
- For developers: the Google Maps Platform Weather API
- For analysts: query the data in BigQuery and Earth Engine
- For bulk use: download from Google Cloud Storage in Zarr format
- For custom runs: a paid inference option on Google Cloud with control over ensemble size and forecast length
- To look at it: the Weather Lab site shows the forecasts in real time
New in this version are forecasts aimed at energy operators: wind speed at 100 metres, roughly the height of a turbine, plus high-resolution cloud cover and solar radiation for estimating output from solar farms.
WeatherNext 3 accuracy and the comparison with WeatherNext 2
Google calls WeatherNext 3 its most accurate global model and points to Brightband’s independent Operational WeatherBench. The public Google launch gives the clearest quantified gains for precipitation. In medium-range forecasts, Google reports CRPS improvements of up to 60% against IMERG, 30% against MRMS and 10% against rain-gauge measurements at early lead times. Those percentages are against different observation baselines, not a single direct 60% improvement over WeatherNext 2.
| Capability | WeatherNext 3 | WeatherNext 2 |
|---|---|---|
| Forecast refresh | Every hour | Every 6 hours |
| Key surface resolution | 5 km | 25 km grid |
| Other surface variables | 10 km | 25 km grid |
| Primary new observations | Live satellite data plus stations | Earlier operational approach |
| Developer data access | BigQuery, Earth Engine, GCS | Open-source weights also available |
| Open-source operational model | No | Yes, weights and code released |
The table highlights an important trade-off. WeatherNext 3 is easier to consume as an operational service, but the current operational model is not open source. Google explicitly directs users who want open code and weights to earlier WeatherNext models, including WeatherNext 2. That suits production data users better than researchers who need to inspect, modify or self-host the exact model.
Lyria 3.5 reaches the Gemini API, but remains a preview
Google’s second model availability change on 3 September was Lyria 3.5 in the Gemini API. Lyria 3.5 Clip generates fixed 30-second clips and costs $0.04 per request. Lyria 3.5 Pro generates structured songs lasting a couple of minutes, supports prompt-controlled duration and costs $0.08 per request. Both take text or image inputs and output 44.1 kHz stereo audio. Pro can also return WAV instead of the default MP3.
The model itself is not entirely new: Google published its model card in July and already used the technology in its own products. The September event is developer availability through the Gemini API. Pricing matches Lyria 3, so the upgrade case is quality and control rather than cost. Google says Lyria 3.5 improves musical coherence, natural vocals and structural control, but does not publish a numerical benchmark that lets buyers quantify the gain over Lyria 3.
There are also preview constraints. Lyria 3.5 has no free tier, preview models can have tighter rate limits, and music generation is single-turn. Developers cannot iteratively edit an existing generation through a sequence of prompts. Google filters requests for specific artist voices and copyrighted lyrics, and every generated track carries a SynthID watermark.
What is uncertain about the WeatherNext 3 rollout
Google’s public launch post is dated 3 September, but its WeatherNext developer release notes date version 3.0 to 31 August and say the 64-member ensemble data was published then. The model and datasets appear to have been staged before the broader public rollout. Teams tracking release dates should distinguish the August 31 developer release from the September 3 public rollout.
Weather forecasts also need local validation. Google’s own article recommends official local weather agencies for warnings. A 5 km global product can be far more useful than a 25 km grid without replacing high-resolution local models or regulatory weather services for safety-critical decisions.
What the WeatherNext 3 and Lyria 3.5 release mean
WeatherNext 3 is worth testing now for renewable-energy forecasting, logistics, agriculture and any location-heavy business that can join forecast data with operational records. Start with one decision where forecast error has a measurable cost, such as wind output, route disruption or staffing by location. Compare the hourly feed against the weather source you use today over several weeks.
Lyria 3.5 is worth testing for creative tooling, but it is not a cost-driven upgrade and remains a preview without multi-turn editing. For most Google customers, WeatherNext 3 is the strategically important change because it turns a research-grade global forecasting model into an hourly data service that can sit directly inside cloud analytics and operational workflows.
For more information, visit the official announcement of WeatherNext 3 on the Google website.
