Google says its AI weather model is getting better
Developed over time first seen 1 week ago
Google has rolled out WeatherNext 3, an updated AI weather forecasting model that it says draws more heavily on live satellite observations rather than relying solely on historical data, allowing forecasts to update every hour instead of every six. The company describes the results as "unprecedentedly" sharp, particularly for predicting rain and snow from fast-moving weather systems, and says the approach helps close forecasting gaps in regions outside the US and Europe that have fewer ground-based rain gauges.
The model produces a global picture five times sharper than its predecessor, WeatherNext 2, visualising variables such as temperature and moisture at up to 5-kilometre resolution compared with the previous 25-kilometre grid, which Google says yields precipitation forecasts up to 50% more accurate a day or more ahead. WeatherNext 3 also forecasts wind speeds at turbine height to support renewable energy planning, and has already been integrated into Google Search, Maps and Gemini, alongside collaborations with agencies including the US National Hurricane Center. Google notes the AI model still trains on and works alongside traditional physics-based forecasting rather than replacing it.
- Google's WeatherNext 3 AI model updates forecasts hourly using live satellite data
- Offers 5x sharper resolution and up to 50% more accurate rain forecasts
- Also predicts wind speeds to aid renewable energy planning
New here? Start with this
Google has released WeatherNext 3, the latest version of an artificial intelligence system it uses to forecast weather. Unlike traditional forecasting, which relies heavily on physics-based models built from historical weather data, this system is trained using machine learning and increasingly draws on live satellite readings to make predictions.
Weather forecasting matters well beyond deciding whether to carry an umbrella: it underpins flood and storm warnings, farming decisions, aviation and shipping safety, and energy planning, particularly for wind power. Google is positioning this technology as a way to make forecasts sharper and more frequent, and is working with official bodies such as the US National Hurricane Center, though it says AI is meant to complement rather than replace conventional forecasting methods.
The development sits within a wider trend of technology companies applying AI to scientific and infrastructure problems, and forms part of Google's efforts to embed its AI tools into everyday products such as Search and Maps.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Advocates of AI-driven forecasting argue that models like WeatherNext 3 represent a genuine public good, particularly for the many regions of the world that lack dense networks of ground-based weather stations and radar. By leaning on satellite data and updating hourly rather than every six hours, such tools could meaningfully improve early warning for fast-moving storms and flooding, potentially saving lives in places traditional meteorology has historically underserved. They also see value in the model's integration with everyday tools like Search and Maps, and its collaboration with agencies such as the US National Hurricane Center, as evidence that AI is complementing rather than replacing established forecasting science.
The case against
Sceptics, including some within the meteorological community, urge caution about placing critical, safety-relevant forecasting increasingly in the hands of a private technology company whose models are not as openly scrutinised or independently verifiable as traditional physics-based systems run by public agencies. They worry about over-reliance on proprietary AI for high-stakes decisions like storm evacuation, and question whether headline accuracy claims, such as being '50% more accurate', fully hold up across all conditions and regions. There is also unease about the optics of a company whose data centres are significant energy consumers marketing tools partly framed around supporting renewable energy planning, and about the broader trend of essential public infrastructure, like weather prediction, becoming concentrated in a handful of large tech firms rather than publicly accountable institutions.
Full account
Google has begun rolling out a new version of its artificial intelligence weather system, WeatherNext 3, which the company says produces sharper and more timely forecasts than its predecessor, particularly for rain and snow. Developed by Google DeepMind, the model is described as offering a global view roughly five times more detailed than the previous WeatherNext 2 system, achieved by drawing on a broader and more current mix of observational data rather than relying mainly on the outputs of conventional physics-based forecasting.
Conventional weather prediction has long depended on supercomputers running complex atmospheric simulations, a process that is computationally heavy and inherently subject to delay. AI-driven models, including Google's earlier systems, have instead learned to spot patterns in historical weather records to generate forecasts more quickly. WeatherNext 3 pushes this further by feeding in live satellite observations alongside data from weather stations, giving it what Google describes as a constantly refreshed picture of atmospheric conditions. As a result, the model can issue new forecasts every hour, compared with the six-hourly updates of WeatherNext 2, and can resolve detail down to a 5-kilometre grid rather than the previous 25-kilometre one. Google says this finer resolution helps particularly with tracking fast-developing storms and improves forecasting in regions with sparse ground-based weather instruments, which tend to lie outside the US and Europe. The company states that precipitation forecasts made at least a day ahead are now up to 50 percent more accurate than before.
Alongside general forecasting, Google has built the new model to support renewable energy planning, generating predictions such as wind speeds at around 100 metres — roughly the height of a turbine — as well as cloud cover and sunlight levels relevant to solar power generation. Google frames this partly in the context of its own rising electricity demands as it expands data centres to support AI development, suggesting that better renewable-energy forecasting is a priority as those demands grow. WeatherNext 3 is set to feed into a range of Google's consumer and developer products, including weather information shown in Search, the Gemini app and Google Maps, as well as the Maps Weather API and Google Earth Engine, and is available for public experimentation through Google's Weather Lab tool.
The two accounts of the launch differ mainly in emphasis. One report leans on direct comment from Google researchers — engineer Samier Merchant and DeepMind scientist Ferran Alet — to explain the shift towards richer, fresher observational data and to frame the renewable-energy angle around Google's broader energy consumption as an AI developer. The other report gives more weight to how the model will be embedded across Google's existing products and tools, and separately notes that an earlier version of WeatherNext was made open source in August 2026, a detail not mentioned in the other account. Both agree on the core technical claims: the jump from a 25-kilometre to a 5-kilometre grid, the move from six-hourly to hourly forecasts, the use of live satellite data to overcome lag in traditional numerical models, and the headline figure of up to 50 percent more accurate precipitation forecasts, with the biggest gains expected in areas that have historically had less reliable forecasting.
Where outlets differ
The Verge-style account attributes the update directly to quoted Google researchers (Samier Merchant on data sourcing, Ferran Alet on renewable energy and Google's own rising energy needs), while the other account presents the claims more as statements from Google DeepMind without named quotes.
The Verge-style account specifies that the biggest forecasting improvements are expected in regions outside the US and Europe, where ground weather stations are sparser; the other account frames the accuracy gains more generally as being in 'historically less reliable' regions without specifying geography.
Only the second account mentions that WeatherNext 3 will power specific existing Google products (Search, Gemini app, Maps, Maps Weather API, Earth Engine) and that the original WeatherNext model was open-sourced in August 2026; the first account does not cover product integration or open-sourcing.
The first account gives more technical framing around the shift from physics-based supercomputer simulation to AI pattern recognition, while the second account focuses more on the practical description of data sources (numerical weather prediction data's six-hour lag versus live satellite feeds).
More coverage