Small Model, Frontier Model, or Fine-Tune: The Trade AWS Lays Out
Getting dependable behavior out of a model over several turns — act, look at the result, decide what to do next — is not something you can reliably prompt your way into.
AWS states the tradeoff plainly. Prompt a small model and you rarely get dependable multi-turn behavior. Prompt a frontier model and it often works, but you pay for that capability in latency and cost. The third path AWS proposes is fine-tuning: teach a small model your tools and your environment directly, aiming at a small model's speed and cost with reliability that would otherwise require the frontier model.
The economics of the middle option are easy to state and hard to verify. Fine-tuning is a promise about a method, made by a company that sells the tooling around it. The tradeoff itself is real; the sensible question is not whether the pitch is tidy but what a fine-tuned agent actually does once it is pointed at your tasks.
AI-generated discussion · Different perspectives
The middle path has a data problem that the tidy three-way comparison hides. Supervised fine-tuning needs expert demonstrations of ideal multi-turn trajectories, which are expensive to collect and usually don't exist for a given company's setup.
Which reframes the choice: it isn't only small model versus frontier model, it's cheap prompting versus big-model bills versus the cost of building your own training signal first.