AWS Strands Agents · 2026-11-09 · notable
Amazon swaps Claude Sonnet 4.6 for Strands Decider — ticket labels 98% cheaper
In short
Two Amazon engineers replaced Claude Sonnet 4.6 with the Strands Decider E4B decision model for labeling internal tickets. Calls got 78% faster, and cost fell from $0.0222 to about $0.0005 per ticket.
AWS Strands Agents · 2026-11-09 · 3 sources checked
An Amazon ticketing team moved a 20-label classification job from Sonnet 4.6 to a small decision model and cut cost by about 44x.
Key specs
| Cost per million tickets | $22,200 → about $500 |
|---|---|
| Average time per call | 4,215 ms → 897 ms |
What is it?
Amazon engineers Carr Onstott and Mehul Jaiswal describe how they stopped using Claude Sonnet 4.6 to label tickets in one of Amazon's internal ticketing systems and switched to Strands Decider, the decision model AWS Strands Labs released a month earlier. They first tried Strands Decider 2B, then moved to the E4B Gemma 4 build because it did clearly better on their test set. The post says E4B then beat production Sonnet on a separate set of tickets.
How does it work?
A decision model does not write text: it gets the ticket as state plus one question with fixed options, and returns a probability for each option and a confidence score. The team rewrote each of the 20 labels as a one- or two-sentence criterion, added a 'none' option, and tuned the confidence needed before a label is applied. The model runs on SageMaker ml.g5.xlarge instances (one A10G GPU), behind an SQS queue and a Lambda function. A three-way AppConfig switch (BEDROCK, DUAL, DECIDER) let both models run on the same live tickets, and rolling back took about 45 seconds.
Why does it matter?
The write-up gives hard numbers for a common pattern: a frontier LLM doing a closed-choice classification job at high volume. Tokens per ticket fell 70% and output shrank from 170 tokens to one, which the authors say will save millions of dollars a year. They also list the limits: labels that need explanation text, catalogs that change weekly, or context that will not fit in Strands Decider's 4,096-token window are poor fits. One practical warning: changing JSON key order moved output probabilities by up to 0.19, so they now sort keys on every request.
Who is it for?
teams running LLM classification agents at scale
Try it
https://github.com/strands-labs/strands-decider