Get early access

AI forecasting7 min read

Can AI predict the weather? WeatherNext 3 and the AI forecasting revolution

Google DeepMind's WeatherNext 3 forecasts the planet every hour at up to 5 km, with up to 50% more accurate rain forecasts. How AI weather forecasting works.

By Kristian FagerlieFounder & CEO, Sikt Intelligence

Short answer: yes, and better than almost anyone expected. The newest example is WeatherNext 3, launched by Google DeepMind and Google Research on September 3, 2026: an AI model that forecasts the weather for the whole planet every hour, down to 5 kilometres for temperature and moisture, with up to 50% more accurate precipitation forecasts a day or more ahead. It caps three years in which AI went from research curiosity to the forecast behind Google Search, Maps and Gemini, and a key tool for hurricane forecasters.

What is WeatherNext 3?

WeatherNext 3 is Google's "most advanced and accurate global weather AI model," built by Google DeepMind and Google Research (Google). What sets it apart:

  • It learns from live observations. Earlier AI models started from the same estimate of the atmosphere that traditional forecasts use, which arrives with a delay of about six hours. WeatherNext 3 also feeds directly on live geostationary satellite images, updated every hour, and is trained on sparse weather-station observations and satellite precipitation data, including NASA's IMERG.
  • Hourly forecasts. It produces a new global forecast every hour, where WeatherNext 2 worked in six-hour steps.
  • Much sharper detail. Temperature and moisture at the surface come at up to 5 km resolution, other surface variables at 10 km and atmospheric variables such as wind at 25 km: roughly five times sharper than before.
  • Better rain forecasts. Google reports up to 50% more accurate precipitation forecasts for planning a day or more ahead. At early lead times, its precipitation error improved by 60% when checked against NASA's IMERG satellite data, 30% against US radar (MRMS) and 10% against rain gauges.
  • Energy and storms. It adds variables for clean energy, such as wind at turbine height and solar radiation, and outputs discrete cyclone tracks.

It powers weather in Google Search, the Gemini app and Google Maps, and developers can use it through the Maps Platform Weather API, Earth Engine, BigQuery and Google Cloud Storage. The research is described in a technical paper (arXiv).

How we got here: the WeatherNext story

  • 2023, GraphCast: Google DeepMind's first big result. It was more accurate than the leading traditional medium-range forecast on 90% of 1,380 test targets, in under a minute on one machine (Lam et al., Science).
  • 2024, GenCast: a probabilistic model that produces many possible futures instead of one. It beat the world's top traditional ensemble forecast on 97.2% of 1,320 targets, for forecasts up to 15 days ahead (Price et al., Nature).
  • 2025, hurricanes in the real world: the US National Hurricane Center used a Google DeepMind cyclone model during the 2025 season. According to NOAA, it was "the most accurate model for storm track and intensity," surpassed only by the Hurricane Center's own official forecasts, and it gave forecasters unusual confidence that Hurricane Melissa would rapidly intensify into a Category 5 storm (ABC News).
  • August 2026, WeatherNext Cyclones in Nature: the cyclone model's three-day forecasts are as good as what earlier models could manage for two days, an extra day of warning that Google DeepMind compares to a decade of meteorological progress. The model, along with WeatherNext 2 and a compact WeatherNext 2-mini, was open-sourced for researchers and forecasters (Google DeepMind; Nature).
  • September 2026, WeatherNext 3: hourly, high-resolution forecasts learned partly from live observations.

How AI weather forecasting works

For decades, forecasts have come from numerical weather prediction: supercomputers solve the equations of the atmosphere step by step, starting from the best estimate of the weather right now. Each run takes hours on some of the largest computers in the world.

AI models learn how the atmosphere evolves from decades of historical weather data instead, then produce a forecast in seconds or minutes. WeatherNext 3 goes a step further by learning from what satellites and weather stations see right now, so the forecast can start from fresher information.

The rest of the field

Google DeepMind is not alone:

  • Pangu-Weather (Huawei, 2023) was the first AI model shown in Nature to beat traditional medium-range forecasts, running about 10,000 times faster (Bi et al., Nature).
  • Aurora (Microsoft, 2025) is a foundation model for the Earth system that outperformed several operational systems (Bodnar et al., Nature).
  • AIFS (ECMWF), from Europe's main forecasting centre, has run in operations since 25 February 2025, side by side with its traditional system, with gains of up to 20% for tropical cyclone tracks on some measures and about 1,000 times less energy per forecast (GeoGarage; ECMWF).

Where AI weather forecasting still falls short

  • Rare extremes. A model trained on history has seen few events beyond the historical range, which is exactly where an unprecedented heat wave or storm sits.
  • The starting point. Most AI models still begin from an estimate of the atmosphere produced by traditional systems; learning from live observations, as WeatherNext 3 does, is a step away from that, not the end of it.
  • Wind and the upper air. Even WeatherNext 3 works at 25 km for atmospheric variables such as wind, so very local effects still need specialised short-range forecasts.
  • Long range. Accuracy falls the further ahead you look, for AI as for physics. Beyond a week or so, a range of outcomes is more honest than a single forecast.

What the weather teaches about forecasting anything

Weather is where the science of grading forecasts began. The Brier score, still the standard way to score probability forecasts, was designed in 1950 to grade weather forecasters, and it is how AI forecasters of elections and world events are judged today. Three lessons carry over:

  • Probabilities beat single calls. "70% chance of rain" is more useful, and more honest, than "rain".
  • Score everything. Weather services publish their verification scores; any forecaster, human or AI, should be judged the same way. Try it on your own forecasts with our Brier score calculator.
  • Fresh evidence matters. WeatherNext 3's biggest gains come from using what is observed right now, and the same is true when forecasting world events.

Real-world events, such as elections or central bank decisions, are harder than weather: there are no equations to learn, and every question is new. We cover how AI handles them in AI forecasting, explained and can AI predict the future?

A live AI forecast, scored in public. In the Sikt Midterm Bench, Sikt forecasts the 2026 midterms next to Kalshi and Polymarket. Latest round, Oct 3, 2026: Senate: Democrats 56% (markets 62% to 64%), every round; House: Democrats 88% (markets 90%), every round. Every forecast will be graded after election day, misses included.

Key takeaways

  • WeatherNext 3 (September 2026) forecasts the whole planet every hour, at up to 5 km for temperature and moisture, with up to 50% more accurate precipitation forecasts.
  • It learns from live satellite images and weather-station observations, not only from traditional forecast data.
  • It powers weather in Google Search, Gemini and Maps, and is available to developers.
  • Google DeepMind's cyclone model was the most accurate hurricane guidance of 2025 after the official forecasts, and in 2026 added about a day of warning.
  • AI still struggles with rare extremes and long range, and weather proves the basics: probabilities and honest scoring.

FAQ

What is WeatherNext 3?

Google DeepMind and Google Research's newest global weather AI model, launched on September 3, 2026. It produces hourly forecasts at up to 5 km resolution for surface temperature and moisture, learns from live satellite data and weather-station observations, and gives up to 50% more accurate precipitation forecasts a day or more ahead.

Where is WeatherNext 3 used?

In Google Search, the Gemini app and Google Maps, and for developers through the Maps Platform Weather API, Earth Engine, BigQuery and Google Cloud Storage.

Can AI predict the weather better than traditional forecasts?

On most measures, yes. GraphCast and GenCast beat the leading traditional forecasts on 90% and 97.2% of test targets, Europe's main forecasting centre has run an AI model in operations since 2025, and WeatherNext 3 adds hourly, high-resolution forecasts learned from live observations.

Can AI predict hurricanes?

It helps a lot. In 2025, according to NOAA, Google DeepMind's cyclone model was the most accurate guidance for storm track and intensity, behind only the National Hurricane Center's official forecasts. Its 2026 version gives three-day forecasts as accurate as earlier models' two-day forecasts.

Sources

Keep reading