Medium-range global weather forecasting · Ceded (since 2023)

GraphCast, published in Science, beats ECMWF's HRES deterministic forecast on about 90 percent of 1,380 verification targets out to ten days; GenCast, in Nature in 2024, beats the ENS ensemble on 97 percent of targets.

What was ceded is skill on the standard headline metrics (RMSE and related scores) at 0.25 degree resolution, measured against the best physics-based system, at a small fraction of the compute per forecast. The comparison is not clean: these models are trained on ERA5, a reanalysis produced by that same physics-based system, and they still depend on its data assimilation for their starting state. ECMWF put its own machine-learned model, AIFS, into operations in February 2025, so the forecasting centres adopted this rather than lost it. Extremes, fine-scale weather and physically consistent fields remain open, and no forecaster's judgement was ever the benchmark; the benchmark was another machine.