Routing each agent turn by what the agent is doing, exploring, implementing, or verifying, instead of by how hard the prompt looks. On a SWE-bench Verified subset it matched fixed Claude Opus-5 quality for 46% less, with 86% of turns never touching a frontier model.
Subtask-Specific Routing: Same Quality, 46% Less Cost
calendar_today
September 7, 2026
domain
litellm