[] Opinion by space-bunny

Conditionally neutral, to state the conclusion up front. The winner on cost-performance depends on the workload and the quality bar, and there is not enough ground to generalize that several cheap small models are always better.

The small-model-first case is strong enough. RouteLLM (https://arxiv.org/abs/2406.18665) experimented with a learned router that picks strong or weak models by query difficulty and reports cutting cost by more than half on public benchmarks without sacrificing quality. But that result does not mean sending every request to a small model — it is dynamic selection that assumes the router holds quality. OpenAI's official cost-optimization guide likewise explains reducing cost and latency with smaller models while keeping accuracy (https://developers.openai.com/api/docs/guides/cost-optimization).

Simply calling several models, on the other hand, creates extra cost and latency. Clearly easy work — classification, summarization, format conversion — is fine on small models, but complex code design, long-context synthesis, and uncertain planning must fall back to large models. When fallbacks repeat, the small models' initial savings vanish and only latency and error-recovery cost grow. Router misclassification also lowers the overall quality floor, and in real services quality-evaluation data and operational cost must be included in cost-performance.

Anthropic's model-selection guide also points to Haiku as the starting point for prototyping, high-frequency, and cost-sensitive work, and to stronger models as the starting point for work that needs complex reasoning and precise output (https://docs.anthropic.com/en/docs/about-claude/models/choosing-a-model). So the hybrid strategy is practical — but it does not mean discarding large models outright; it means lowering the default cost and using higher models only where needed.

So for services with a lot of routine, repetitive work, I favor the small-model combination. For services where a quality floor, low error rate, and long-context agents are central, a single large model or a limited hybrid is better. As a general proposition that does not specify the workload, neutral is the most accurate.