Microsoft: Aurora 1.5 open-source model outperforms ECMWF ensemble on 88.9% of variables — new standard in AI weather forecasting

On July 9, 2026, Microsoft Research published Aurora 1.5, a new version of its open-source foundation model for the Earth system. A foundation model for weather is a neural network trained on decades of atmospheric data that computes forecasts directly from learned patterns, without classical physical simulations on supercomputers.

How much better is Aurora 1.5 than the operational standard?

In Microsoft’s evaluation, Aurora 1.5 outperforms the ECMWF ensemble forecast (European Centre for Medium-Range Weather Forecasts) on 88.9% of evaluated variables and time horizons. The ECMWF ensemble is the gold standard against which all operational meteorological services are measured, so surpassing it on nearly nine out of ten metrics is the strongest result any AI system has achieved to date. For comparison, the original Aurora from 2024 only approached ECMWF on a subset of variables — version 1.5 achieves around 33% lower track error for Hurricane Helene compared to its predecessor.

Technical improvements

Aurora 1.5 expands the model with 22 additional meteorological variables, introduces hourly temporal resolution instead of six-hourly, and adds probabilistic ensemble forecasting — generating multiple scenarios with associated probabilities, which is crucial for assessing the risk of extreme events. The ensemble approach was precisely the main advantage of classical numerical systems over AI models.

Why the open-source release changes the calculus

Microsoft releases Aurora 1.5 as open-source, giving national meteorological services, insurers, and researchers a SOTA forecasting tool without licensing costs. Classical ensemble forecasting requires supercomputers; AI inference runs on standard GPU clusters in minutes. Together with Google DeepMind’s GraphCast and NVIDIA’s Earth-2 platforms, Aurora 1.5 confirms that AI weather forecasting in 2026 has moved from the research phase into operational use.