It compiled, and it called the wrong server.

Ask a model for the code that calls a web service and it returns something that imports cleanly and builds a well-formed request 81% to 92% of the time. Working from the request alone, the strongest model gets the host, path and method all correct 3.76% of the time, and the wrong host is 69.8% of every failure. Hand over the service’s own spec and the same models reach 88.7% to 90.5%.

Wang, Fan, Zhu, Li and Wang · 2026 · 3,135 real call functions · 850 services · 2,082 repositories · seven models