If a model’s only job is to choose a label, paying it to write JSON may cost more than the decision itself.
In a November 9 case study, Amazon reports that replacing Claude Sonnet 4.6 with Strands Decider cut the model cost of an internal ticket workflow from about $22,200 to $500 per million tickets. The figures use published AWS prices and exclude Lambda. The reported average model call fell from 4,215 to 897 milliseconds.
What Amazon changed
Amazon used a Strands Decider build based on Gemma 4 E4B. The team shortened the prompt and reduced each of its 20 labels to a short criterion. It added a “none” option. The replacement produces a choice instead of a JSON answer. For long tickets, Amazon drops the oldest comments to fit the model’s context window.
The team first kept Sonnet in charge and ran the two models on the same tickets. It used their disagreements to check whether the new labels were good enough. The team also found that changing the order of JSON keys shifted the output probabilities. It now sorts keys before sending requests.
Calculate the cost at your own traffic level
At Amazon’s cited rate of $1.408 per instance hour and roughly 5,600 tickets per hour at full use, the instance alone costs about $251 per million tickets. At half that throughput, the same instance costs about $503 per million. Low or uneven traffic can change the comparison before you add fallback and operating costs.
For a trial, compare both classifiers against tickets people have labeled. Inspect rare categories separately and send low-confidence cases to a reviewer or your current model. Count the cost of that fallback before switching. For workflows that also need to generate answers or carry context between steps, see our guide to evaluating Strands Harness.
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
