June 18, 2026
Your theory is right on both counts. The memory wasnt durable -- it was living in the provider session, so swapping gateways wiped it. Keep it in a file or db the agent reads each run and a model switch wont touch it.
The cost blowup is OpenRouter auto-routing. Leave it on auto and it grabs pricey models. Pin explicit model IDs instead -- cheap default for the bulk (planning, classification, routine edits), and only send the design work to Claude.
Youre not married to GPT. You just need to pin per task instead of letting the gateway pick. Delegate by task type, cheap by default, escalate only where the quality actually changes the output.