Amazon engineers swap Claude Sonnet for their own AI, slashing ticket-labeling costs by 98%

Two Amazon engineers, Carr and Mehul, working on the company's massive internal ticketing system, replaced Anthropic's Claude Sonnet 4.6 LLM with Strands Decider, a specialized decision model Amazon launched a month ago. The old system used Sonnet to label millions of tickets—customer escalations, capacity requests, security findings—by generating JSON with a long system prompt. The new approach uses Strands Decider, which doesn't generate text but instead takes the ticket as input and returns probabilities for each possible label from a fixed catalog, requiring far less computational power.

The swap to the smaller, more focused E4B Gemma 4 build of Strands Decider performed just as accurately as Sonnet on production tickets but at 98% lower cost, saving Amazon millions in annual token expenses. The engineers deployed it on SageMaker instances, with the entire pipeline using standard AWS services like SQS, Lambda, and CloudWatch. For developers, this is a powerful case study in optimizing AI costs: replacing a general-purpose LLM with a targeted, efficient model for high-volume, repetitive tasks can yield massive savings without sacrificing performance.