Hurricane Isaias landed where the AI guidance pointed. That still grades nothing

Hurricane Isaias came ashore near Destin, Florida, at about 8 p.m. Eastern on October 9 as a Category 2 storm, on the eastern side of a two-way split in the guidance that forecasters were watching all week. According to WDSU meteorologist Derek Sibley, the National Hurricane Center had been leaning toward the AI models, Google's among them, which mainly favored the eastern side of the track, while most traditional physics-based models sat on the western side, with plots over Mississippi and Louisiana. The official forecast aligned with the AI family, and so did the storm. It is the clearest live test AI weather guidance has had in front of a public that did not ask for one, and it grades nothing by itself.

Two families of guidance, and one disagreement worth naming

Sibley's framing describes how the job is done now, not how it is advertised. Meteorologists draw on two families of guidance: the traditional physics-based models, which simulate the atmosphere from physical equations, and newer AI versions, which learn the relationship between a given atmospheric state and what happens next. For Isaias he said the National Hurricane Center "has been leaning more toward the AI models including Google AI, which has mainly favored the eastern side of the forecast," while "most of the traditional physics based models have been leaning more toward the western side of the forecast... with plots over Mississippi and Louisiana."

The WDSU comparison is therefore not a story about a model existing. It is a story about two families of guidance pointing at different stretches of coastline, and an official forecast that picked one. CBS News put landfall at Destin, in the Florida Panhandle, which sits on that eastern side. That outcome is suggestive, not scored. No verification of the October forecasts has been published, and the forecast a storm lands on the near side of is still a single event, not a skill measurement.

What the storm itself did

Metric	Reading
First hurricane of the 2026 Atlantic season	Formed Oct 7, the latest first hurricane in 121 years (NOAA data via CBS News)
Peak strength	Category 3 briefly on Oct 9
Landfall	Destin, Florida, about 8 p.m. ET on Oct 9, at Category 2
Winds	90 mph sustained at landfall; a 112 mph gust at Santa Rosa Sound
Power	More than 697,000 customers out across Florida, Alabama and Georgia
Emergency status	States of emergency in Florida, Alabama and Georgia; federal declarations for Florida and Alabama

The season had been unusually quiet, widely attributed to a strong El Nino, part of why the October landfall drew so much attention. None of that speaks to which model family was right.

The one published scorecard is from last season

The only skill comparison anyone has actually published is the 2025 season, and it is more interesting for its shape than its headline. WDSU reports that Google's DeepMind model outperformed the traditional American and European models on track, and was the best-performing model especially between 12 and 96 hours, or up to four days out. It beat even the National Hurricane Center's official forecasts on that window, except on day five. On intensity the picture narrows sharply: the Google model and the NHC were nearly neck and neck, with Google ahead only between 36 and 60 hours. A large track edge and a thin intensity edge is a specific claim, and it is the kind of claim a single storm cannot confirm or refute.

WDSU also states that 2025 was the first hurricane season in which the NHC incorporated AI model guidance into its forecasts, and that the agency expects AI models to become "more reliable, efficient, and influential components of the hurricane forecasting process in the coming years." Both are about the blend, not about the model winning.

The vendor's own case, dated as background

Google DeepMind's WeatherNext team published its cyclone work in Nature and released the code and weights on August 6, well before Isaias existed, so it belongs to the background rather than to this month's news. The claims are substantial and vendor-authored. The model predicts a cyclone's track, intensity and wind structure with state-of-the-art accuracy, and on average gains, in DeepMind's wording, "more than a full day (24 hours) of lead time advantage" — its three-day forecasts match what prior models delivered at two days, which the team describes as roughly a decade's worth of meteorological progress. It runs 1,000-member ensembles, up from 50 a year earlier, to sample rare tail risks such as rapid intensification; a 15-day forecast generates in under a minute on a TPU; and it trains on nearly 20 terabytes of global atmospheric data plus the IBTrACS archive of nearly 5,000 historical storms. The cyclone model needs only 28x28 km input resolution, which DeepMind says is 100 times coarser than traditional models and remains an open research question. The team worked with the NHC, CIRA and the UK Met Office, and says the system helped the NHC in 2025 track Hurricane Melissa's rapid intensification and its Jamaica landfall.

The extra day is the line most likely to be repeated, and it should be read as the vendor's comparison against prior models, not an operational gain measured outside DeepMind. USA Today's framing of the same tension is worth holding next to it: for decades, track accuracy improved slowly and intensity improvement lagged further behind, so people "could go to bed at night expecting a Category 2 storm the next day and wake up to a Category 5."

What would actually settle the question

Three things, none of them available yet. The NHC's verification report for the 2026 season, with per-storm error tables, would put track and intensity errors for AI and physics guidance side by side on the same storms. Whether agencies keep AI guidance in the blend after a season in which it lands on the wrong side of the track would say more about institutional confidence than any single forecast. And a larger sample of storms than one that happened to come ashore where the AI plots pointed. The 2025 numbers are one season and one agency's comparison, with a wide track edge and a narrow intensity edge that both bend with lead time. Isaias is a data point a public happened to watch in real time, and it is not yet evidence of anything except that the guidance is now being used.