Intel / article

Amazon saves millions by replacing Sonnet 4.6 with Strands Decider

Source: strandsagents.com
Author: aloukissas
Date: Published Nov 10, 2026

Context

Two Amazon engineers who work on an internal ticketing system write that they replaced Anthropic's Claude Sonnet 4.6 with Strands Decider for one narrow job: putting each incoming ticket, such as a customer escalation, capacity request, automated alarm or security finding, into one of 20 categories. Strands Decider is a decision model from the Strands Agents project. It cannot write text; given a situation and a set of options, it returns a probability for each option and a confidence score. Earlier Intel explains the original 2B version. This team used an E4B build based on Google's Gemma 4, which did clearly better on their test set.

The switch required reshaping the task. Sonnet had read a long for every ticket and written JSON with its reasoning. The team turned labelling into a closed choice, rewrote each label as a one- or two-sentence criterion, added a "none" option, and tuned how confident the model had to be before a label was applied. The model runs on SageMaker instances with one 24 GB A10G GPU each, and the endpoint scales with traffic. Its 4,096-token forced shorter label descriptions and means the oldest comments on long tickets are dropped. The team also found that reordering keys in the same JSON input shifted the model's probabilities by up to 0.19, so it now sends keys sorted.

All figures are the authors' own, and the costs are computed from published list prices. Average time per call fell from 4,215 ms to 897 ms, and tokens per ticket from 6,724 to 2,010, with output dropping from 170 tokens to 1. Cost per ticket fell from about $0.0222 to about $0.0005, or about $22,200 against $500 per million tickets, which they describe as about 44 times cheaper; the $500 includes spare capacity for bursts. The "millions" in the title is their projection of yearly token savings, with no total given. On quality, they ran both models side by side on live traffic and judged every disagreement, and they say Decider performed just as well and beat production Sonnet on a separate set of tickets, but they publish no accuracy figures. They expect the approach to suit tasks with a stable list of options and short inputs, not ones that need explanation text or more context than a few thousand tokens.

Terms in this piece · Glossary

system prompt — The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source strandsagents.com