Open SWE model router reportedly cut median cost per thread by 64%, with no measurable quality change

LangChain says an A/B test across 973 Open SWE coding-agent threads compared its model router with always using its strongest model. Median cost per thread was $0.94 with routing, versus $2.61 without it—a 64% reduction. LangChain reports no measurable change in quality: 29.2% of routed threads produced merged pull requests, compared with 27.3% of control threads. The router chooses a model at the start of each thread and uses it throughout.