Microsoft's Aurora 1.5 beats ECMWF on 88.9% of forecasts

Aurora 1.5 beat the European Centre's ensemble forecast on 88.9% of targets, and Microsoft released it with open weights, per the company's July 10 research post.

That's a real shift for anyone forecasting without access to a proprietary ensemble. Code sits on GitHub and checkpoints sit on Hugging Face. A lab that could never afford ECMWF-grade access can run a model that beats it.

The new version adds 22 weather variables and hourly temporal resolution on top of the earlier open Aurora model. It also adds probabilistic ensemble forecasting, which the earlier version didn't have.

On tropical cyclones, including Hurricane Helene, the ensemble median cut track error by about a third against the same ECMWF baseline.

The 88.9% figure matters less than the license. Anyone can run a model that beats one of the world's top forecasting centers without paying for proprietary access. I'd watch whether smaller forecasters start running Aurora 1.5 directly instead of waiting on ECMWF's own updates.