Amazon cuts ticket-labeling costs 98% by switching from Sonnet to Strands Decider

Amazon replaced Claude Sonnet 4.6 with Strands Decider, a decision model, for classifying tickets across their internal system. While matching Sonnet’s performance, Strands Decider reduces costs by 98%—from generating text and requiring long prompts to returning probability scores for a fixed label set. The E4B Gemma 4 build runs on SageMaker instances, processing requests in roughly 640 milliseconds with 19 GiB memory. In practice, Strands Decider costs around $500 per million tickets, making it 44 times cheaper than Sonnet. Lambda execution duration dropped 74% after the switchover.

The team validated the switch through parallel testing: a three-way configuration ran both models on real traffic to compare results before full migration. This real-world validation proved more informative than offline testing and enabled simple rollback if needed. The approach succeeded because the task involves fixed categories (20 labels), high volume, and latency sensitivity. The engineers cautioned that decision models cannot replace language models when tasks require explanatory text, frequently changing catalogs, or ticket context that exceeds the 4,096-token window.