Google DeepMind and Google Research introduced the new AI model WeatherNext 3 on September 3, 2026, which monitors atmospheric changes in greater detail and forecasts the weather more frequently. Google said that data obtained from the model would be used in search results, Google Maps and Gemini, and would also be made available to users and researchers on Google’s cloud platforms.
WeatherNext 3 demonstrated the highest accuracy among leading models in metrics such as temperature, wind speed and humidity in Brightband’s Operational WeatherBench tests, which compare AI forecasts. The model outperformed the deep-learning systems of Google, Microsoft, Nvidia and the European Centre for Medium-Range Weather Forecasts (ECMWF), as well as the traditional forecasts of the US National Weather Service and ECMWF.
Traditional forecasts are prepared by processing mathematical equations that describe the physics of weather on supercomputers operated by government agencies. Although these systems produce accurate results, they are expensive and relatively slow. After ECMWF released more than half a century of weather data in 2018, researchers began developing deep-learning models capable of achieving similar accuracy more quickly.
WeatherNext 3 targets issues such as AI forecasts focusing on broad areas, performing poorly on precipitation and relying on formatted datasets from government agencies. The model can forecast some key variables at a resolution of 5 km; precipitation assessments improve by 60% compared with WeatherNext 2, and it can produce hourly forecasts instead of forecasts every six hours. Containing 2.4 times more parameters than the previous model, the system was trained to forecast not only hurricane tracks but also the conditions that specific weather stations could measure.
Because the model can process satellite data collected in real time on an hourly basis, it can produce forecasts more frequently. While Google says WeatherNext 3 is the “first” AI model to use raw observations directly in high-resolution global forecasts, it states that WindBorne WeatherMesh 6 has been using raw observations since late 2025. Both models rely on national weather datasets.
The low cost and speed of AI forecasts could provide economic benefits in poorer regions where accurate weather information is inaccessible because of expensive sensors and supercomputers. More detailed wind, precipitation and cloud forecasts are also expected to make renewable energy projects more reliable.
Why it matters
This development is reinforcing the shift in weather forecasting from physics-based systems that rely on the expensive supercomputers of government agencies to AI models that can process raw observations more quickly. Bringing hourly and higher-resolution results to Search, Maps and Gemini means that advances in forecasting infrastructure are entering the daily flow of information for a broad user base directly. Cloud access could lower the barrier to use for researchers and regions that lack access to expensive sensors and computing infrastructure; wind, precipitation and cloud data are also highly relevant to renewable energy projects. However, the model’s reliance on national weather datasets, the debate over the claim that it is the first system to use raw observations, and the extent to which its performance in benchmark tests will translate to real-time conditions remain open questions.
Background
Google is not a new name in the FikirPilot archive: we have published 17 news reports mentioning the name in the past 90 days; the most recent is dated 4 September 2026.