WeatherNext 3 forecast accuracy now leads every major competitor, according to Brightband’s Operational WeatherBench, as Google DeepMind and Google Research release the most capable version yet of their AI-driven weather system.
The model bests deep-learning rivals from Microsoft, Nvidia, and the European Centre for Medium-Range Weather Forecasting (ECMWF), and outperforms traditional physics-based forecasts from the US National Weather Service and the ECMWF. Google says WeatherNext 3 will feed into what users see on Google Search, Maps, and Gemini, and will be available via Google’s cloud platforms.
‘This is going to be the first time that some of the core variables feed and power a lot of the Google products,’ Samier Merchant, a Google senior staff engineer, told TechCrunch.
Sharper Grids, Hourly Updates, Better Rain
The headline improvement is resolution. According to the Google DeepMind WeatherNext 3 announcement, the model renders key surface variables such as temperature and moisture at a 5-kilometre (0.05°) grid, other surface variables at 10 kilometres, and atmospheric variables such as wind speed at 25 kilometres. Its predecessor, WeatherNext 2, produced forecasts on a 25-kilometre (0.25°) grid in six-hour increments, making the new model roughly five times sharper on its finest variables.
The timing has improved too. WeatherNext 3 is initialised every hour using live geostationary satellite input, enabling hourly forecasts rather than the six-hourly standard. The WeatherNext 3 research paper (Rasp et al., 2026) notes that hourly output covers all single-level variables, including solar radiation and cloud cover, which were not available in WeatherNext 2.
Rain forecasting has historically been an Achilles heel for AI weather models. WeatherNext 3 trains against multiple precipitation datasets, including ECMWF reanalysis, NASA’s IMERG satellite data, and Google’s own satellite-radar information. Google’s research benchmarks report reductions of up to 50% in certain precipitation forecast error scores compared with numerical weather prediction baselines under specific evaluation conditions. The snippet figure from researchers is a 60% improvement on rain over WeatherNext 2 directly.
The model is also larger: 2.4 times more parameters than its predecessor, with decoder heads tuned to provide more practically useful outputs, including predictions targeted at specific ground-level weather stations rather than averages across a three-dimensional grid.
‘The idea, with a lot of AI applications, is to try to run tasks as end-to-end as possible,’ said Daniel Rothenberg, an atmospheric scientist at Brightband. ‘Adding a capability where this model is now also predicting, say, what Denver’s airport’s weather station is going to measure on an hourly basis, just connects that forecasting task closer to the core.’
WeatherNext 3 Forecast Accuracy in Context: A Field in Motion
The backdrop is a discipline undergoing rapid structural change. When Brightband launched the Operational WeatherBench dashboard to compare AI and physics-based forecasts, WeatherNext 2 was already the world’s top medium-range model, narrowly ahead of ECMWF’s AIFS-ENS. Brightband notes that, in the current generation, four of the top five models on its leaderboard are AI-based rather than physics-based.
The shift traces back to 2018, when the ECMWF released more than half a century of archived weather data. Deep-learning researchers began training models that could generate forecasts far more quickly and at a fraction of the cost of the government supercomputers that had dominated the field. Ferran Alet, a staff research scientist manager at DeepMind, put the underlying logic plainly: ‘Weather is chaotic, and so small differences really start to perturb massively… Machine learning targets the problem we are really solving, which is approximate noisy physics from incomplete information and finite compute, and so it learns patterns from a lot of data.’
Google’s path through this transformation has been eventful. Its older GraphCast model (now called WeatherNext Graph), a deterministic graph neural network, outperformed the ECMWF’s HRES in over 90% of tested cases and variables, was named a runner-up to Science’s Breakthrough of the Year in 2023, and won the MacRobert Award in 2024.
WeatherNext 3 marks a further step: it is an ensemble model (rather than deterministic), and Google claims it is the first AI model to directly incorporate raw observations for a high-resolution global forecast. WindBorne, an AI weather start-up, contests that framing, saying its WeatherMesh 6 model has incorporated raw observations from its weather balloon fleet and other sources since late 2025. Google’s counter is that its forecasts are higher resolution globally. Both models still rely on national weather datasets, so full direct data assimilation remains unfinished work. The Operational WeatherBench currently tracks models from NOAA, ECMWF, Google, Microsoft, and Nvidia, with further AI entrants due to be added.
The stakes extend well beyond consumer apps. Alet pointed to higher-resolution forecasts of wind, rain, and cloud cover making renewable energy projects more dependable. Bill Gates has cited AI-powered weather forecasting as a crucial benefit of the technology, with better forecasts improving crop yields in developing countries. The low compute cost also promises to extend high-quality forecasting to regions where government supercomputers have simply been too expensive to deploy.
‘At the end of the day, I think Google is about providing useful information to the user, and a lot of what users are looking for has to do with the weather in some way or another,’ Alet said.
The immediate test will be whether WeatherNext 3 holds its lead on Brightband’s leaderboard as WindBorne, NOAA, and others publish their next-generation models in the months ahead.
