Google DeepMind Performance During the 2026 Hurricane Season: AI Models Discussion 11-12-2026

With Google DeepMind consistently used in the National Hurricane Center (NHC) forecast process now, we thought it would be worthwhile to evaluate its performance during the 2026 season. This blog will focus on some bulk statistics, and we will later take a look at some individual cases. We’ll start with the East Pacific, which was much busier in 2026 due to the record El Niño event.

East Pacific

Figure 1 shows the track forecast errors for the East Pacific for 2026 for a few models: operational GFS, operational HAFS-A, operational HAFS-B, the "HFIP Corrected Consensus" (HCCA, which tends to be one of the key pieces of consensus guidance), and Google DeepMind. The NHC forecast is also shown. The results are striking: Google DeepMind outperformed individual models, and was even slightly better than the NHC forecast at some lead times.

Another way to look at this is “frequency of superior performance” (FSP), which is the frequency with which a given model had the best forecast. For all lead times (Figure 2), Google DeepMind had the highest FSP for track forecasts in 2026. It was particularly dominant in the 72-120 hour time frame.

Looking at intensity errors, the results are similar (Figure 3). Google DeepMind was the best-performing individual model at all lead times, and comparable to the NHC forecast for most lead times as well.

The one aspect where Google DeepMind did show some room for improvement was the intensity bias (Figure 4). Despite the better overall performance, GDMI did show a fairly consistent negative bias, even more so than HAFS-A at longer lead times. The NHC forecast (and HCCA consensus) was much better for this metric. This will be something that is hopefully improved in future versions of the model.

Overall, this was an extremely impressive performance by Google DeepMind during a record-breaking season, highlighting that AI models really are helping to lead the way in improving TC forecasts.

Atlantic

The Atlantic was much quieter during the 2026 season, but it is still worthwhile to examine the model performance. Figure 5 shows the track forecast errors for the Atlantic from the same set of models, as well as the NHC forecast. The difference between the models was not quite as pronounced as it was in the East Pacific, but Google DeepMind was similarly the best-performing model at most lead times.

Figure 6 shows the frequency of superior performance for the Atlantic. Google DeepMind was by far the best performer at earlier lead times (hours 24-48). At longer lead times, NHC and GFS were comparable, but the sample size was also much lower, due to the lack of long-tracked systems in the Atlantic this year.

For intensity errors, the story is fairly similar (Figure 7). Google DeepMind was the best individual performer at Days 1-2. HAFS was comparable or better at 72 hours, but the small sample makes it difficult to draw robust conclusions.

For the intensity bias (Figure 8), the results are fairly similar to the East Pacific. Google DeepMind has a negative intensity bias that is slightly worse than HAFS early on, though the overall bias is generally similar for most of the models across lead times.

Overall, the performance was encouraging despite the very small sample during the quiet Atlantic 2026 season. There is definitely room to improve, particularly with the intensity bias, and hopefully we will see those improvements in future versions of Google DeepMind and other AI models.