LLM routers must recover cache costs before switching models
For coding agents, choosing the cheapest capable model is not enough: switching models can discard a conversation's warm cache and make the next model rewrite the full history. Edgee frames the decision as ski rental, balancing that one-time rewrite against the savings from cheaper future turns. Its Agent Gateway simulates 30 days of sessions weekly, tests tier maps and switching rules against hindsight, and recommends a lower-cost combination under guardrails.