Amazon and AWS have been using Anthropic’s balanced model, “Claude Sonnet 4.6,” to automatically route internal inquiries and system errors with AI while keeping costs down. They have reported success in reducing costs by 98% by replacing it with the decision-making model “Strands Decider.”

Amazon saves millions by replacing Sonnet 4.6 with Strands Decider | Strands Agents
https://strandsagents.com/blog/amazon-saves-millions-replacing-sonnet-with-strands-decider/

“Strands Agents” is an open-source toolkit developed by AWS for building AI agents for production environments. An explanatory article about Amazon’s migration from Claude Sonnet 4.6 to Strands Decider has been published on the official Strands Agents blog.


Amazon’s internal ticketing system categorizes every type of internal event, including customer escalations, capacity requests, automated alerts, and security issues, into 20 categories. Amazon had been using Claude Sonnet 4.6 as the AI responsible for this categorization. However, because categorization does not involve generating text or images, Amazon determined that a decision-making model that outputs probabilities and confidence scores for each option, rather than a high-performance large language model, would be optimal, and replaced it with its proprietary decision-making model, “Strands Decider.”

Although Amazon initially used the standard Strands Decider, it developed “Strands Decider E4B,” a four-billion-parameter model based on Gemma 4, in pursuit of further performance improvements. By adjusting the level of confidence the model must have before assigning a label, it achieved performance that surpassed Claude Sonnet 4.6 in production.

Claude Sonnet 4.6 outputs responses in JSON format, whereas Strands Decider E4B outputs only the selected option. This makes prompts far shorter than when using Claude Sonnet 4.6 and dramatically reduces the number of input and output tokens needed to process each ticket.

◆Input tokens per ticket
Claude Sonnet4.6: 6554
Strands Decider E4B: 2009

◆Output tokens per ticket
Claude Sonnet4.6: 170
Strands Decider E4B: 1

The following graph shows the execution time of AWS’s serverless computing service, Lambda, at the time of the migration from Claude Sonnet 4.6 to Strands Decider E4B. The migration to Strands Decider E4B reportedly reduced Lambda’s average execution time by 74%.


This reduced the cost of processing one million tickets from $22,200 (approximately ¥3.5 million) to $500 (approximately ¥79,000). This represents a 98% cost reduction compared with Claude Sonnet 4.6.

Furthermore, Strands Agents states: "When a label requires a description, when the catalog changes weekly, or when useful context cannot be condensed into a few thousand tokens, a decision-making model may not be able to replace the work of an LLM."