WeatherNext matched leading cyclone forecasts a day earlier

What one extra day actually means

A cyclone forecast does not become useful at one fixed moment. Track, strength and the reach of dangerous winds all sharpen at different rates as a storm develops.

A peer-reviewed paper published in Nature on 6 August reports that WeatherNext Cyclones reached the accuracy of leading operational models about a day earlier on average. The comparison covered tropical cyclones from 2023 through 2025 and looked at track, intensity and wind radii.

That is a lead-time advantage, not a promise that every storm will be called correctly one day sooner. Some forecasts will gain more, some less, and a useful model output still has to be read alongside observations and other guidance.

Even so, a day is meaningful in this field. It can give human forecasters more time to examine a possible turn, a rapid strengthening event or a wider wind field before official advice goes out.

A thousand plausible futures

WeatherNext Cyclones is an ensemble model. Instead of drawing one future path, it can generate many plausible versions of the atmosphere and the storm inside it, extending as far as 15 days.

The system was trained on global atmospheric analysis data and the International Best Track Archive for Climate Stewardship, a record covering nearly 5,000 historical storms. That specialist cyclone history helps it follow more than the centre line. It also predicts intensity and the size of the wind field.

Google says a single forecast takes less than a minute on one of its tensor processing units. The system can scale from the 50 members common in conventional ensembles to as many as 1,000, giving rare but serious possibilities more chances to appear in the spread.

More members do not make uncertainty disappear. They make it easier to see its shape.

The surprising part is the coarse view

Specialist regional models often use very fine grids because a cyclone's inner structure matters. WeatherNext works with atmospheric cells roughly 28 kilometres across, according to Google's technical description.

The Nature paper says that, despite those coarser inputs, the model's intensity and wind-radius forecasts were competitive with leading operational systems. The authors argue that global atmospheric data contains more signal about cyclone strength than researchers had assumed.

They also tested the model inside a weighted consensus forecast. Adding WeatherNext improved that combined guidance, which is a more practical result than asking one system to replace everything else.

The finding does not show that resolution no longer matters. Local rainfall, storm surge and the details near land still demand other tools. It shows that a global model can contribute more to cyclone guidance than its grid size might suggest.

Open code, with a fairly large asterisk

Google has published WeatherNext code, pretrained weights and sample data. The code is under Apache 2.0, while other repository materials use a Creative Commons licence.

There is also a smaller model that can run in a free Colab session. Google says the full versions are optimised for its TPUs and need an H100-class GPU when run on graphics hardware. The smaller version is not expected to match the paper's main results.

The repository calls this research code, provided as-is, with no promise of a stable interface. Reproducing a forecast is therefore more accessible than training the whole system or operating it reliably every day.

Most importantly, the model is not a warning service. Google and the paper's weather-agency collaborators present it as guidance for expert forecasters, not something the public should use instead of an official alert.

What is confirmed, claimed and still open

Confirmed: Nature published the peer-reviewed study on 6 August 2026. Its author list includes researchers from Google, the US National Hurricane Center, Colorado State University and the UK Met Office. Google has also released the model code and weights.

The research finding: in the authors' 2023-2025 evaluation, WeatherNext Cyclones offered about a day or more of average lead-time advantage across track, intensity and wind-size forecasts. Ensembles of up to 1,000 members represented rare outcomes better than smaller ensembles in the tests reported.

Google's claim: the company describes the result as roughly a decade of operational progress delivered at once. That comparison is based on the historical rate at which forecast errors have fallen; it is an interpretation, not ten years of live deployment.

Still open: how the model performs through future seasons, unusual basins and changing climate conditions; how much forecasters gain from it in practice; and where higher-resolution regional models remain essential. The public release makes independent testing possible, but the largest versions still require substantial hardware.

This is a strong result with a sensible role. A new forecast joins the desk. The warning still comes from people responsible for issuing one.