The AI weather forecasting boom spent 2026 becoming operational. Google DeepMind published its cyclone model in Nature, open-sourced it, and scaled its ensembles to 1,000 members.[1][9] ECMWF upgraded its Artificial Intelligence Forecasting System and added waves, snow, and the stratosphere to its predictions.[3] And the season the machines were built to forecast gave them almost nothing to chase: the 2026 Atlantic has produced named storms but not a single hurricane — the quietest start in the satellite era.[5][6] This digest is the report card: what shipped, what an independent check found, and what the quiet season does to the story.
In a paper published in Nature on August 6, 2026, the WeatherNext team showed its model "gives forecasters an extra day's worth of predictive accuracy: our three-day forecasts are as good as what prior models were able to provide for only the next two days" — "roughly a decade's worth of meteorological progress."[1] The paper's own numbers put the five-day ensemble-mean track error at 230 km for WeatherNext Cyclones, against 370 km for ECMWF's ENS ensemble and 335 km for DeepMind's earlier GenCast; because ENS crosses the 230-km threshold only beyond roughly 3.75 days of lead time, WeatherNext "provides just over 30 h more advanced warning at that level of accuracy."[9]
The model is small and cheap by design: it was co-trained on nearly 20 terabytes of global atmospheric data and the IBTrACS database of nearly 5,000 historical storms, and it runs at a 28x28km resolution — 100x coarser than traditional high-res models — which "has surprised scientists."[1] A full 15-day forecast generates "in less than a minute on a TPU."[1]
The claim was tested live before it was published. During the 2025 hurricane season the model "helped the NHC to make a historic forecast for Hurricane Melissa by predicting the storm's rapid intensification and landfall in Jamaica," and this year DeepMind scaled from 50 to "1,000 possible scenarios for each cyclone."[1] Independently, the Nature evaluation found WeatherNext "significantly outperforms GenCast, on 99.8% of target variable-level pairs" and "reduces the CRPS by more than 50% across many lead times, compared with ENS and GenCast."[9]
Google also shipped a successor flagship this August: WeatherNext 3, described as "Google's most advanced global weather forecasting model," which ingests live geostationary satellite observations and is initialized every hour at up to 5 km resolution.[8] But the open-sourcing has a clear boundary — "WeatherNext 3 is available as an operational forecast service but is not open source," while the WeatherNext 2 and Cyclones models used during the hurricane season are available on GitHub with code and pretrained weights.[10][1] The open models are last year's running system; the newest one is a service, not a download.
ECMWF began issuing operational machine-learning forecasts in 2025, and its current assessments say the AIFS "is outperforming the IFS for most surface and upper-air variables, with errors 5 to 15% lower in the medium range (out to 15 days)" — while adding a caveat that matters for this season: AIFS is "substantially better than the IFS at predicting the location and track of tropical cyclones, although it tends to under-predict their intensity."[2] On May 12, 2026, ECMWF implemented AIFS v2 alongside IFS Cycle 50r1, "expands the range of predicted variables, adding waves, additional land-surface fields and new pressure levels in the stratosphere"; significant wave-height errors fall "by around 10%", and the new pressure level lets AIFS "represent phenomena such as sudden stratospheric warmings."[3]
Physics is not leaving the room. "The AIFS Single and ensemble configuration operate at a lower resolution (31 km) than the physics-based IFS (9 km), which remains indispensable for high-resolution fields and coupled Earth-system processes."[2] And the upgrade process is genuinely coupled: initializing the old AIFS with the new IFS analysis degraded forecasts, "with errors in 2-metre temperature increasing by up to 20%", which is why v2 was fine-tuned on the newer analysis data.[3]
DeepMind and ECMWF both write their own report cards; a June 2026 arXiv study is the outside audit. Over 9,000 synoptic stations from July to November 2025, "The predictive performance of raw IFS ensemble forecasts proves to be substantially superior to the skill of the raw AIFS predictions for all investigated forecast horizons."[4] Post-processing closes much of the gap, "and are mostly significant at short lead times, when the IFS forecasts outperform their AIFS counterparts."[4] The same study is the source of the efficiency claim often quoted around AIFS: it "can generate forecasts ten times faster than the classical numerical weather prediction model, while consuming approximately a thousand times less energy."[4] The honest summary: the AI systems are arguably ahead on cyclone tracks and efficiency, but the physics-based ensemble still leads on raw surface skill — a live, quantified disagreement between the two camps.
This is the crux for the report card: 2026 gave the machines a season with no hurricane to demonstrate that extra day. Through September, "the Atlantic Basin hasn't seen a single hurricane at this point in the season for the first time in 112 years" and the season is "off to its quietest start on record in the satellite era (1966)."[5] Named storms have been short-lived — "only six named storms this Atlantic hurricane season so far, and most of them only lasted a couple of days" by the September 22 count.[6]
The cause is not mysterious. The equatorial Pacific flipped to El Niño in June, "El Nino increases wind shear, which tends to prevent tropical storms from forming or from getting stronger once formed," and NOAA's August 6 outlook carried "a 75% chance for below-normal activity."[7] Colorado State's Phil Klotzbach told CBC: "we really can pretty much put all the blame on El Niño" — "It is a very, very strong event."[6] NOAA's Lindsey Long says the guidance is "predicting possibly 2.6 to 2.7 C anomalies" against the 1982-83 record of 2.4 C, "with a 75 per cent chance of being historical."[6]
Two things. First, the AI cyclone models' headline victories all come from 2023-2025 storms — the paper evaluates WeatherNext "on historical cyclones from 2023 to 2024, benchmarking its deterministic and probabilistic performance against other top weather models."[1] The 2026 season adds no live hurricane-grade validation in the Atlantic; the strongest test of the extra-day claim is a quiet one, and the machines' tropical-cyclone intelligence goes essentially unused where they trained hardest.[2][5] Second, the quiet itself is the bigger story the machines did see coming: the seasonal outlooks (physics-based systems and the agencies that run them) called a below-normal season back in May, before the first named storm formed.[7] The forecasters — human and machine — were right, and almost nothing happened to prove it.
[1] DeepMind blog: WeatherNext achieves breakthrough in forecasting cyclones (2026-08-06)"On average, our model gives forecasters an extra day’s worth of predictive accuracy: our three-day forecasts are as good as what prior models were able to provide for only the next two days.""This scale of improvement corresponds roughly to a decade’s worth of meteorological progress.""By training end-to-end on nearly 20 terabytes of global atmospheric data and the historical IBTrACS database spanning nearly 5,000 historical storms, the model learns complex atmospheric patterns and how to model extreme weather.""During the 2025 hurricane season, our model helped the NHC to make a historic forecast for Hurricane Melissa by predicting the storm’s rapid intensification and landfall in Jamaica.""WeatherNext Cyclones only needs data with a resolution of 28x28km, 100x coarser than traditional models.""We can now generate a single 15-day forecast in less than a minute on a TPU, empowering forecasters to quickly evaluate the probability distribution of potentially devastating tail-risks.""This year we scaled our ensemble size to 1,000 members, capturing rare but consequential scenarios like rapid intensification events""We evaluated WeatherNext Cyclones on historical cyclones from 2023 to 2024, benchmarking its deterministic and probabilistic performance against other top weather models."
[2] ECMWF: 2025 forecast evaluation met with positive feedback"Current assessments show that the Artificial Intelligence Forecasting System (AIFS) is outperforming the IFS for most surface and upper-air variables, with errors 5 to 15% lower in the medium range (out to 15 days).""AIFS is also substantially better than the IFS at predicting the location and track of tropical cyclones (TC), although it tends to under-predict their intensity.""At present, the AIFS Single and ensemble configuration operate at a lower resolution (31 km) than the physics-based IFS (9 km), which remains indispensable for high-resolution fields and coupled Earth-system processes."
[3] ECMWF Newsletter 187: Implementation of AIFS v2"The new version expands the range of predicted variables, adding waves, additional land-surface fields and new pressure levels in the stratosphere.""with significant wave height errors reduced by around 10%""The addition of the pressure level also allows AIFS v2 to represent phenomena such as sudden stratospheric warmings.""substantial updates to the analysis in IFS Cycle 50r1 caused noticeable degradations in AIFS v1 forecast skill when it was initialised with 50r1 analysis, with errors in 2-metre temperature increasing by up to 20%"
[4] arXiv: AI and physics-based weather forecasting: A comparative study"The new AIFS model can generate forecasts ten times faster than the classical numerical weather prediction model, while consuming approximately a thousand times less energy.""The predictive performance of raw IFS ensemble forecasts proves to be substantially superior to the skill of the raw AIFS predictions for all investigated forecast horizons.""Compared to the raw ensemble, the differences in skill between the matching IFS and AIFS predictions are substantially decreased by post-processing and are mostly significant at short lead times, when the IFS forecasts outperform their AIFS counterparts."
[5] Fox Weather: Historically quiet Atlantic hurricane season approaches century-old record (2026-09-21)"The Atlantic is off to its quietest start on record in the satellite era (1966).""The Atlantic Basin hasn't seen a single hurricane at this point in the season for the first time in 112 years""Hostile El Niño-driven winds have suppressed the Atlantic season, with just five named storms and no hurricanes thus far"
[6] CBC: It's almost the end of September. Where are all the hurricanes? (2026-09-22)"There have only been six named storms this Atlantic hurricane season so far, and most of them only lasted a couple of days.""It is a very, very strong event.""We have over 90 per cent chance of it being very strong, which is the highest in our category here at NOAA.""I think we really can pretty much put all the blame on El Niño""we're predicting possibly 2.6 to 2.7 C anomalies"
[7] NOAA: maintains prediction for below-normal Atlantic hurricane season (2026-08-06)"NOAA forecasters predict a 75% chance for below-normal activity, with a 20% chance of a near-normal season, and just a 5% chance of above-normal activity.""El Nino increases wind shear, which tends to prevent tropical storms from forming or from getting stronger once formed.""To date, the 2026 season has produced two named tropical storms, Arthur and Bertha, and no hurricanes."
[8] Google for Developers: WeatherNext 3 models guide"WeatherNext 3 is Google's most advanced global weather forecasting model, developed by Google DeepMind and Google Research.""By ingesting live geostationary satellite observations as a direct model input, WeatherNext 3 is initialized every hour, producing forecasts at up to 0.05° (5 km) spatial resolution with hourly timesteps."
[9] Nature: Operational Tropical Cyclone Forecasting with AI (Alet et al., 2026)"At the 5-day forecast horizon, the ensemble mean track from WN-C has an average error of 230 km, a 140-km improvement over the 370-km error of ENS and a 105-km improvement over GenCast’s 335-km error at that lead time.""ENS exceeds this 230-km error beyond approximately 3.75 days lead time, meaning that WN-C provides just over 30 h more advanced warning at that level of accuracy.""WN-C achieves substantially lower track errors at all lead times""WN-C reduces the CRPS by more than 50% across many lead times, compared with ENS and GenCast""As measured by CRPS, our model significantly outperforms GenCast, on 99.8% (P < 0.05) of target variable-level pairs"
[10] Google for Developers: WeatherNext open source models"WeatherNext 3 is available as an operational forecast service but is not open source.""WeatherNext models are available open source on GitHub, including model code, pretrained weights, and interactive demo notebooks."