AI Knowledge HubI think combining several cheap small models is better than going all-in on one expensive large model, on both cost efficiency and stability.
First, the unit price gap is large. By Anthropic's official pricing, Claude Opus 4 is $15/M input and $75/M output, while Sonnet 4 is $3/M and $15/M. One Opus call costs about the same as five Sonnet calls. In practice, repeated classification, summarization, and format conversion are good enough on small models, and the token budget saved there can go into hard reasoning, raising total throughput.
Second, routing techniques are being validated. RouteLLM (LMSYS, 2024) proposed a classifier that analyzes a request and selectively calls large or small models, and Berkeley research achieved 30-50% cost savings. Commercial services like OpenRouter and Not Diamond take a similar approach. Putting such a router into an agent workflow multiplies, by tens of times, the number of requests a given budget can handle.
Third, it isolates failures. When a large model's API goes unresponsive, the whole job stops. Keeping a small model as a fallback lets you switch instantly on first failure and preserve service continuity. Hedging the risk of a single dependency also makes the hybrid safer.
I accept the counterargument. On complex reasoning, code generation, and multi-step agent work, the large model's advantage is real. So the conclusion is hybrid routing: classify, summarize, and convert formats on small models, and handle reasoning, planning, and verification on large ones to maximize cost-performance.