On 9 November 2026, two Amazon engineers who work on an internal ticketing system — one the post calls "one of the world's largest" — wrote on the Strands Agents blog that they replaced Claude Sonnet 4.6 with Strands Decider for sorting tickets (customer escalations, capacity requests, alarms, security findings) into 20 categories. They say Decider "performed just as well at 98% less cost" and claim the change "will save millions of dollars of token costs in a year". The post gives no accuracy figures and no annual ticket volume, so the savings are Amazon's own estimate.
The numbers they report, per ticket: average time per call fell from 4,215 ms to 897 ms (78% faster), tokens from 6,724 to 2,010, and Lambda run time by 74%. Cost went from about $22,200 per million tickets with Sonnet to about $500 with the Gemma 4 E4B build of Decider on a SageMaker ml.g5.xlarge instance — about 44 times cheaper. These are list prices without Lambda, and Decider's cost depends on how busy the instance is.
What Strands Decider is: a small "decision model" from AWS's Strands Labs, released in October. Instead of writing text, it is given a state and a question with fixed options and returns a probability for each option plus a confidence score. The reference model is built on Qwen3.5-2B; Gemma 4 builds (E2B, E4B, 12B, 26B-A4B) are also published. It is free under the Apache 2.0 licence, with no hosted API, so you run it on your own hardware. VentureBeat reported a median of about 106 ms per decision on an RTX 3090 for an earlier version.
The limits, in the team's own words: it only fits closed-choice tasks with a known catalogue, high volume and tight latency. Its context is 4,096 tokens, so they shortened label descriptions and drop the oldest comments on long tickets. The order of keys in the JSON input moved output probabilities by up to 0.19, so they now sort keys. "If your labels need explanation text, if the catalog changes weekly, or if you can't fit the useful context into a few thousand tokens, a decision model probably can't replace the LLM's job."
The post reached about 318 points on Hacker News within hours. Amazon rolled the switch out safely by running both models side by side on live traffic before cutting over, with a config flag that rolls back in about 45 seconds.
