Mixing AI models only pays off in two shapes

If you're pairing an expensive AI model with a cheap one to save money, Anthropic's cost guide has a sentence worth reading twice: in their measurements, "a second model paid off in two shapes".

Shape one, the advisor: the low-cost model runs the whole loop and phones a frontier model only when stuck. Advice comes back short, roughly 400–700 tokens, so you pay top rates only for judgment, never for the bulk generation. Their data says the advisor closed 50–90% of the capability gap to the stronger model — with a catch I'll get to.

Shape two, the orchestrator: the frontier model plans and delegates, cheap workers absorb the token-heavy exploration, and most of the bill lands at worker rates. This is the shape we run daily, though we arrived at it for context-window reasons and the billing benefit came along free.

The catch in shape one deserves a name: the executor has to know it's stuck, and that self-knowledge is itself a capability. In their measurements, a pairing that consulted happily at default effort dropped to almost never consulting when the executor's effort was lowered — and then scored below the executor running alone. You saved money on the worker, and the worker stopped raising its hand.

Two more honest notes from the same guide: prompt caching remains the big lever (2.5–3.7× in their benchmarks; model mixing is "narrower"), and pairing two near-equal models bought a point or two that "a single run does not separate from noise".

Their advice on where to start is almost anticlimactic, and we'd co-sign: sweep effort on your current model first. Most workloads end there.