The Pulse

Amazon says Strands Decider replaced Sonnet for ticket labeling

Amazon engineers say they switched an internal ticket-labeling system from Claude Sonnet 4.6 to Strands Decider E4B. They report similar labeling performance at 98% lower cost, using published prices and their estimate of typical instance u

AI.info Team · Written with AI · fact-checked · 2026-11-09

A different model for a fixed choice

Amazon engineers say they replaced Claude Sonnet 4.6 with Strands Decider E4B for labeling tickets in Amazon’s internal ticketing system, reporting similar performance at 98% lower cost. The Strands Agents account, dated Nov. 9, describes the change as a switch from a large language model to a decision model.

Amazon’s teams had used Sonnet to sort customer escalations, capacity requests, automated alarms, security findings and other company tickets into 20 categories. The previous setup sent each ticket with a long system prompt and asked the model to reason over fixed labels and return JSON. Strands describes Decider as a model that returns a probability for each available option and a confidence score, rather than generating text.

The engineers first tested Strands Decider 2B, then switched to an E4B Gemma 4 build after it performed better on a test set. They say that, after they tuned how confident the model needed to be before applying a label, E4B beat production Sonnet on a separate set of tickets. The team shortened the descriptions of the 20 labels into one- or two-sentence criteria and added a “none” option to the choice list.

Testing the change on live traffic

For serving, the team used SageMaker ml.g5.xlarge instances, each with one A10G and 24 GB of memory; the model used about 19 GiB. The post puts request time at about 640 milliseconds on one instance and says SageMaker copies the model weights from S3 before startup. The endpoint scales with traffic, while its network is isolated from outbound access to the container.

The ticketing pipeline uses SQS, Lambda, a SageMaker AI Endpoint, AppConfig and CloudWatch, with deployment stages named Beta, Gamma and Prod. The engineers used a three-way AppConfig setting to control the transition: BEDROCK ran Sonnet alone, DUAL ran both models, and DECIDER ran Strands Decider alone. During DUAL operation, Sonnet’s labels were applied to tickets while its results were compared with Decider’s.

The post reports average call times of 4,215 milliseconds for Sonnet and 897 milliseconds for Strands Decider E4B, a 78% improvement. It gives total token use as 6,724 per ticket for Sonnet and 2,010 for Decider, a 70% reduction. The team says Decider’s shorter prompt and one-choice output account for the lower token count, while long tickets may lose their oldest comments to fit the model’s context. The post also reports a 74% drop in average Lambda execution duration on the day Sonnet use was turned off.

Costs and limits

Using published us-east-1 on-demand prices, the engineers estimate Sonnet at $0.0222 per ticket and Decider at about $0.0005 per ticket. For one million tickets, their figures are about $22,200 for Sonnet and about $500 for Decider, with the latter depending on how busy each instance is. They say the estimate excludes Lambda costs and assumes headroom for bursts; at full use, one instance could handle about 5,600 tickets an hour for $250 per million tickets.

The engineers found that reordering JSON keys could move output probabilities by up to 0.19, so they now sort keys in every request. They say the approach suited a high-volume task with a fixed label catalog, but may not fit cases requiring explanatory text, a catalog that changes weekly, or more context than a few thousand tokens. The team says running both models on real traffic helped it assess disagreements, and that reverting the switch required a 45-second configuration change.