Not Diamond just released the methodology behind their model routing which gets Opus xhigh quality while cutting agent costs 20–80%.

Model routing becomes much harder decision for long-running coding agents. So Not Diamond is benchmarking the router across simulated user turns, changing task complexity, and response delays that can expire the KV cache.

And harder than it seems, switch models too aggressively and you can throw away the KV cache. Pick one model from the opening prompt and you are assuming task complexity stays fixed for the rest of the session.

So Not Diamond is treating routing as a sequential decision problem instead.

At each step, its router predicts future reward and cost for a particular model and reasoning effort, using current and previous session state, message and token counts, task complexity, KV-cache state, and intermediate reward signals.

Read more detail on their technical report.