Case study from the Strands Agents team. An Amazon ticket-labeling pipeline swapped Sonnet 4.6 for Strands Decider, a small open decision model that only picks between options, and cut cost per ticket from $0.0222 to about $0.0005.
Strands Decider is a decision model, not an LLM: it gets a state and a question with options, and returns a probability for each option plus a confidence score. It can't generate text. The repo is Apache-2.0 and ships a strands-decider CLI and server you install with pip.
In the case study, each ticket is the state and the 20 labels become one-or-two-sentence criteria in a single choice question, plus a "none" option. Labels apply only above a tuned confidence threshold. The rollout ran through a three-way config switch: Sonnet only, both in parallel with results compared, then the decision model only. They report average latency down from 4,215 ms to 897 ms, output tokens from 170 to 1, and roughly 44 times lower cost per ticket.
The broader point: a lot of LLM calls inside agent workflows are really classification. When the answer is one of a fixed set of options, a small model with a calibrated confidence is faster and cheaper, and the confidence tells you when to fall back to a person.