June 16, 2026
Heres what I run daily: Opus does the hard reasoning and planning, then Codex comes in as an independent second set of eyes to cross-check the diff before anything lands. Local models only get the cheap high-volume stuff that doesnt need to be right the first time -- classification, bulk summarization.
The thing I learned the hard way -- a smart router that "figures out what to send where" sounds great but usually costs more than it saves. A dumb default-to-the-strongest-model rule, with a manual drop to local for bulk, beat every clever router I tried. Watch where your tokens actually go by task type first, then route. Dont over-engineer it before you have that data.