A fleet of small, task-focused AI models may cost 10–30× less than one giant model doing everything.

This reads the question as an economics-of-architecture question: is a fleet of small heterogeneous models cheaper *and* competitive with a single large homogeneous pool? The corpus leans toward yes — but with a sharp condition attached about *when* the selection has to happen. The strongest economic case comes from agentic work: most of what an agent actually does is repetitive, well-scoped language tasks, and small models handle those at 10–30× lower cost, which makes 'small by default, large only when needed' the rational design rather than a compromise Can small language models handle most agent tasks?. The heterogeneity itself is the lever, not a tax you pay for saving money.

Routing is where the economics get vivid. Instead of one frontier model answering everything, you send each query to whichever specialist is best for its semantic neighborhood — and a pool of ten 7B models routed this way beat GPT-4.1 and 4.5, while a routed system matched GPT-5-medium at 27% lower cost Can routing beat building one better model?. The claim underneath is that *selection is a stronger lever than scale*: you get more from choosing the right small model than from making one model bigger. That connects to a quieter finding — small models can absorb frontier-level reasoning from good post-training rather than raw parameter count, so a 3B model matches much larger systems on verifiable tasks Can small models match frontier reasoning without massive scale?, and DPO on a teacher's good/bad examples lets small models hit high accuracy on function-calling Can small models match large models on function calling?.

There's an even cheaper flavor of heterogeneity: don't ship many models, ship one base plus millions of lightweight adapters. Treated as persistent behavioral state, PEFT adapters let a single strong base carry per-user or per-task deltas, replacing millions of full models — the economic dream of personalization at population scale Can lightweight adapters replace millions of personalized models?. And on a per-sample budget, smaller isn't just cheaper — models around 500M actually generate *more* unique outputs, because big models concentrate probability mass and give you less variety per dollar Why aren't bigger models better for generating diverse outputs?.

The important catch: cheap small-model pools only match strong models when there's a way to *tell which answer is right*. A committee of weak calls amplifies coverage — the correct answer is probably in there somewhere — but sampling alone can't select it; you need an external soundness signal like a test, proof, or type check to convert latent-correct into actually-chosen When can weak models match strong model performance?. This is also why the frontier-from-post-training result is explicitly bounded to *verifiable* tasks with checkable ground truth. So the economic win is real but domain-shaped: it's strongest exactly where correctness is machine-checkable (code, math, function calls, agent subtasks) and weakest where it isn't.

One more reframe worth carrying away: 'small vs large' isn't the only axis — you can also spend at inference instead of at training. Test-time compute trades off against parameter scaling, letting smaller models match larger ones on hard prompts Can inference compute replace scaling up model size?. So the real economic picture is a three-way budget — parameters, post-training, and inference compute are substitutable resources, and the heterogeneous-small architecture wins by refusing to pay for scale it can buy more cheaply elsewhere.