Picking one of 20 ticket categories does not always require generating an explanation. Amazon engineers report replacing Sonnet 4.6 with Strands Decider E4B and reducing cost from $0.0222 to about $0.0005 per ticket. The migration changed more than the model.
The November 9 first-party account separates measured latency in “What we measured” from list-price estimates in “What it costs.” These are company results; we have not reproduced them.
The comparison includes preprocessing
Sonnet read a long system prompt and generated JSON and reasoning. Decider receives state, a question, and options, then returns option probabilities and confidence. The engineers tried 2B first and selected the E4B Gemma 4 build after better ticket results. This deployment differs from the model in our Decider 2B introduction.
They shortened each label to one or two sentences, added a “none” option, and dropped the oldest comments from long tickets to fit a 4,096-token window. The evaluated unit is therefore the model, label definitions, and preprocessing together. Tickets whose earliest comments contain decisive evidence deserve their own evaluation slice.
An instance-hour is not a token price
| Metric | Sonnet 4.6 | Decider E4B |
|---|---|---|
| Average call time | 4,215ms | 897ms |
| Total tokens per ticket | 6,724 | 2,010 |
| Cost per ticket | $0.0222 | About $0.0005 |
The account compares Bedrock’s global endpoint with SageMaker ml.g5.xlarge. Cost estimates use published us-east-1 on-demand prices and exclude Lambda. The SageMaker instance has one A10G with 24GB and costs $1.408 an hour. A separately reported request time of roughly 640ms informs capacity planning; it is not the same observation as the 897ms average call time.
An editorial calculation makes the utilization assumption explicit: instance cost per ticket is 1.408/q dollars when q tickets are processed per running instance-hour. Reaching $0.0005 requires approximately 2,816 tickets per instance-hour. A lightly used dedicated GPU will have different economics. The 98% reduction is not a universal classifier discount.
Confidence changes coverage
The engineers tuned the threshold required to apply a label. A higher threshold can reduce automatic coverage. The account provides no numerical accuracy score, evaluation sample size, or threshold, so readers cannot independently calculate quality equivalence. Track accepted-label accuracy alongside automatic coverage and review cost.
The team also reports probability shifts of up to 0.19 when JSON keys changed order. It now sorts keys. Reproducible comparisons need stable serialization, label wording, truncation, and model versions.
Copy the rollout contract
BEDROCK calls only Sonnet. DUAL calls both models while applying Sonnet’s result. DECIDER uses only the replacement. This lets operators inspect disagreements before changing the applied label. DUAL also incurs both inference paths’ costs during evaluation.
A stable taxonomy, short useful context, and enough traffic make this a reasonable experiment. Frequently changing labels or mandatory explanations weaken the fit. Include a conventional classifier where appropriate, and compare the final cost of a correctly resolved ticket, including review and omissions, rather than the inference invoice alone.