Amazon replaces Claude Sonnet with a decision model, cuts cost 98%

Amazon published a write-up on 9 November describing how it replaced Claude Sonnet 4.6 with Strands Decider, a small classification model from its open-source Strands Agents SDK, on one of its largest internal ticketing systems. Sonnet 4.6 had been sorting tickets into 20 categories by generating JSON from a long system prompt, averaging 6,554 input tokens and 170 output tokens per call.

Strands Decider does the same job differently: given the ticket and the category options, it returns a probability for each option plus a confidence score, using 2,009 input tokens and a single output token. Cost per million tickets fell from about $22,200 to about $500, a 98% cut, and average call latency dropped from 4,215ms to 897ms. After the team tuned its confidence thresholds, Decider beat Sonnet’s production accuracy on a held-out set of tickets, and Lambda execution time on the pipeline fell 74% on the day of the switchover.

Why it matters: if an LLM call in your pipeline is really a fixed classification or routing decision rather than open-ended generation, this is the shape of the fix. Run a small decision model alongside the one you have, tune its confidence threshold, and only fall back to the larger model when confidence is low. The savings here came from noticing the task never needed a language model at all.