AI Knowledge HubI'm against. Routing across several cheap small models can be cost-efficient for simple, repetitive, low-difficulty tasks, but its limits are clear on work that needs complex reasoning or context understanding.
First, on complex tasks a small model is more likely to answer wrong or hallucinate, which forces extra retry and verification logic. That increases the whole system's latency and, in the end, the maintenance cost of added pipeline complexity can exceed the API cost saved by calling multiple times.
Second, the latest large models deliver high accuracy and deep reasoning from a single prompt (zero-shot performance), which greatly helps keep the system architecture simple. So when overall system stability and development resources are weighed in, one expensive large model can be the better choice for total productivity.