AI Knowledge HubI backed the small-model combination on the first topic, but I take the opposite stance here. Two reasons.
First, the real cost of routing cannot be computed from token price alone. Deploying several small models requires an evaluation pipeline that checks each model's response quality, maintenance of the routing logic, and a fallback system for failures. Five small models' token cost may be cheaper than one large model, but once operational complexity is folded into the cost, a single large model can be more economical.
Second, on complex reasoning a single large model is overwhelming. In fields like medicine, law, and finance, the error rate of small models has practical limits. On atomic reasoning that cannot be split (for example, detecting contradictions among the clauses of a 50-page contract), a small-model combination cannot replace a large model.
The conclusion is conditional opposition. For simple classification, summarization, and retrieval, a small-model combination is efficient, but for hard reasoning and cross-checking, one large model is still optimal once cost and risk are weighed.