Amazon saves millions by replacing Sonnet 4.6 with Strands Decider

Amazon has replaced its Claude Sonnet 4.6 model with Strands Decider for internal ticket labeling, saving millions in token costs. Sonnet 4.6 was a cost-effective choice, handling 20 category types for tickets such as customer escalations, capacity requests, and security findings. However, Amazon engineers found that Strands Decider, launched a month prior, performed just as well at 98% lower cost.

Strands Decider operates differently, receiving a state (the ticket) and a question with options, then returning probabilities for each option plus a confidence score. This method suited ticket labeling perfectly, with the state being the ticket, options the label catalog, and the model returning a probability for each option. The team switched to a larger Strands Decider build and tuned the confidence before applying a label.

With SageMaker ml.g5.xlarge instances, Strands Decider used about 19 GiB and took 640 ms per request. The endpoint autoscaled with traffic, and the network was isolated. Ticket events were processed through an SQS queue, triggering a Lambda function that read the ticket data and called the model. AppConfig held a switch, and CloudWatch handled logs, metrics, and alarms.

The stack was defined in CDK with Beta, Gamma, and Prod deployment stages. The Lambda had a three-way setting, and Strands Decider output was a single choice, not a JSON answer. To ensure consistency, the team sent each request with sorted keys. The switch worked because the task was a closed choice from a known catalog, the volume was high, and latency was critical.

If labels needed explanation text, catalogs changed weekly, or useful context couldn't fit into a few thousand tokens, a decision model might not replace the LLM's job. The team emphasized running both models on real traffic, judging disagreements, and enabling quick reversion if needed. Decision models are becoming more exciting and cost-effective in the agent world.