Open SWE's model router reportedly cut median thread costs 64% in a 973-thread test

LangChain reports that an A/B test of its Open SWE coding agent compared routing each thread to one of three models with always using its strongest model. Median LLM cost per thread was $0.94 with routing, versus $2.61 for the control—a 64% reduction. Merged-PR rates were 29.2% and 27.3%, respectively, a difference LangChain says was not statistically significant. The router chose a model at the start of each thread.