What AI can now do·Nov 14, 2023·Science
GraphCast beats the top physics forecast on 90% of 1,380 targets
90% of 1,380Targets where GraphCast beat ECMWF HRES
GraphCast (Science) was more accurate than ECMWF's HRES on more than 90% of 1,380 variables and lead times, producing a 10-day global forecast in under a minute on one TPU v4 machine.
How far it has travelled
Demonstratedshown once, by its developer or a researcher
yes · Nov 14, 2023 · Science paper; >90% of 1,380 targets and 99.7% in the troposphere; a live run predicted Hurricane Lee's Nova Scotia landfall about 9 days out versus about 6 for traditional forecasts. Huawei's Pangu-Weather (Nature, 2023) had already reported roughly 10% lower RMSE than ECMWF's IFS on reanalysis data. source
Reproducedshown again by someone else
yes · Feb 25, 2025 · Pangu-Weather (Huawei) reported ~10% lower RMSE than IFS and cyclone track errors of 120 km at 3 days versus 162 km for HRES; ECMWF then built its own AIFS, found up to 20% more accurate over 18 months of testing. source
Accessibleavailable to the public or to customers
yes · Nov 14, 2023 · GraphCast code open-sourced on GitHub; ECMWF published live GraphCast forecasts on its website. source
Reliableworks most of the time on a defined task
yes · Feb 25, 2025 · ECMWF's AIFS became operational on Feb 25, 2025 after 18 months of testing: up to 20% more accurate on some measures, 12 hours more lead time on tropical-cyclone tracks, about 1,000x less energy. source
Real consequenceshas changed something in the world, for good or ill
yes · Feb 25, 2025 · AI forecasts now feed the public weather services of ECMWF's 35 member states; experts flag remaining weaknesses in predicting storm intensity. source
What this could change for you
The forecast on your phone increasingly comes from a model trained on past weather rather than a physics simulation, with longer warning before hurricanes.
Sources
primary source, second source.
A milestone is a dated, sourced result that moved what AI can do; a stage is upgraded only when a source shows it (a reproduction, a public release, production evidence, a documented consequence). “Unknown” is the honest default.
