tl;dr: Amazon replaced Claude Sonnet 4.6 with Strands Decider E4B (a Gemma-based decision model that outputs probabilities over fixed options rather than text) for labeling internal tickets, cutting costs ~98% (~$22,200 to ~$500 per million tickets) while improving accuracy and latency by 78%. The team ran both models in parallel via an AppConfig switch to validate before cutover. The approach works when classification is a closed-choice problem with a stable catalog and context that fits in a few thousand tokens; one gotcha was that JSON key order affected output probabilities by up to 0.19.