Amazon and AWS teams have replaced Claude Sonnet 4.6 with Strands Decider for their internal ticketing system, achieving significant cost savings while maintaining performance. The decision model, launched a month ago, performed just as well as Sonnet 4.6 at 98% less cost, potentially saving millions of dollars annually in token expenses. The system processes tickets across 20 category types, including customer escalations, capacity requests, automated alarms, and security findings.
The transition involved tuning model confidence thresholds and leveraging the E4B Gemma 4 build for optimal results. Running on Sage Maker ml.g5.xlarge instances, Strands Decider processes requests in approximately 640 milliseconds. The implementation utilizes AWS App Config for three-way settings between Bedrock, dual-model comparison, and the Decider mode.
Lambda functions read ticket data from SQS queues and call the model endpoint, with CloudWatch monitoring metrics and alarms. The architecture is defined in CDK with deployment stages for Beta, Gamma, and Prod environments. Benchmarks show Strands Decider reduces average time per call by 78% and total tokens per ticket by 70% compared to Sonnet 4.6.
The token efficiency stems from producing a single choice output instead of JSON responses and utilizing shorter prompts. This optimization demonstrates how decision models can effectively replace general-purpose LLMs for structured classification tasks within large enterprise systems.
Source: Hacker News · Summarized by HeadlinesBriefing