--- title: Amazon saves millions by replacing Sonnet 4.6 with... description: Amazon engineers replaced Claude Sonnet 4.6 with Strands Decider, a decision model, to label tickets in a large internal ticketing system, cutting costs by... canonical: https://daily.dev/posts/amazon-saves-millions-by-replacing-sonnet-4-6-with-strands-decider-bfmmo8l36 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Amazon saves millions by replacing Sonnet 4.6 with Strands Decider | daily.dev og:description: Amazon engineers replaced Claude Sonnet 4.6 with Strands Decider, a decision model, to label tickets in a large internal ticketing system, cutting costs by... og:url: https://daily.dev/posts/amazon-saves-millions-by-replacing-sonnet-4-6-with-strands-decider-bfmmo8l36 og:image: https://api.daily.dev/og/posts/BfMmo8l36.png og:image:alt: Amazon saves millions by replacing Sonnet 4.6 with Strands Decider og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Amazon saves millions by replacing Sonnet 4.6 with Strands Decider **[Hacker News](https://daily.dev/sources/hn)** · 5 min read · 0 upvotes · 0 comments ## Summary Amazon engineers replaced Claude Sonnet 4.6 with Strands Decider, a decision model, to label tickets in a large internal ticketing system, cutting costs by 98%. Strands Decider E4B performed just as well as Sonnet while running on SageMaker ml.g5.xlarge instances at about 640 ms per request. Average time per call dropped from 4,215 ms to 897 ms, a 78% improvement, and tokens per ticket fell from 6,724 to 2,010. Cost per million tickets dropped from about $22,200 to about $500. The team also found that JSON key order could shift output probabilities by up to 0.19, so they now sort keys before sending requests. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much cheaper is Strands Decider than Claude Sonnet 4.6 for ticket classification? Strands Decider E4B costs about $500 per million tickets versus about $22,200 for Claude Sonnet 4.6, a 98% reduction, roughly 44 times cheaper. Sonnet's cost comes from processing 6,724 tokens per ticket at $3 per million input and $15 per million output tokens, while Strands Decider uses about 2,010 tokens per ticket on a SageMaker ml.g5.xlarge instance at $1.408 per instance-hour. _Comparing LLM versus decision-model costs for a classification pipeline is easier with real benchmarks tracked on daily.dev._ ### What is a decision model and how does it differ from an LLM for classification tasks? A decision model can't generate text; given a state and a question with a set of options, it returns a probability for each option plus a confidence score, rather than generating JSON or free text. This fits closed-choice tasks like ticket labeling well, where the state is the ticket and the options are a fixed label catalog, cutting both latency and token usage versus an LLM. _Developers weighing decision models against LLMs for classification can follow comparisons like this on daily.dev._ ### Does the order of JSON keys in a prompt affect a decision model's output? Yes, sending identical content with JSON keys in a different order moved Strands Decider's output probabilities by up to 0.19. The fix was to always send requests with sorted keys. Anyone serving a decision model behind an API should check for this sensitivity before comparing runs or benchmarking results. _Catching subtle serving quirks like key-order sensitivity matters when daily.dev readers deploy decision models in production._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 82 comments across hackernews (as of 2026-11-10). **TL;DR:** The discussion centers on debating when LLMs are overkill for classification tasks versus when their zero-shot flexibility justifies the cost, with many pointing out this pattern (prototype with an LLM, then swap in a cheap specialized classifier) is a recurring and sensible lifecycle rather than a novel insight. **Sentiment:** 20% positive · 55% mixed · 25% skeptical **The case for** - LLMs let teams bootstrap a classifier in minutes without labeled data or a data scientist, which is valuable even if later replaced. - The LLM-to-specialized-model pipeline is seen as a legitimate and increasingly common pattern for de-risking prototypes before committing to training a leaner model. - Some argue running an expensive model briefly to generate training/synthetic data, then switching to a cheap classifier, is a reasonable cost tradeoff at scale. **The pushback** - Several commenters call using an LLM as a classifier wasteful hype-driven engineering when simpler, faster, cheaper classical methods (SVM, BERT, scikit-learn) would suffice. - Some question why they didn't just try cheaper existing options like Haiku 5.5 or other small/Chinese models instead of building custom infrastructure. - The post is viewed by some as thinly veiled marketing for Amazon's own Strands Decider product rather than a neutral case study. - One commenter is confused by the article's internal inconsistency about which model (Strands Decider vs. Gemma) was actually used. - The sensitivity of outputs to JSON key order is flagged as concerning, with some noting this is well-known behavior that should have been anticipated. **By community** - hackernews (mixed): Debate splits between those who see this as an obvious, sensible return to efficient classifiers and those who find the premise annoying marketing or confusing, alongside tangents about ticket culture and why cheaper existing models weren't tried first. **Hottest debate:** Whether using LLMs for classification tasks is ever justified versus always being wasteful 'AI grift' when simpler classical ML would do the job. **Open questions** - Why wasn't a cheaper off-the-shelf model like Haiku 5.5 or an existing decision model (Jev/Luna) tried before building this custom pipeline? - Is the inconsistency between 'Strands Decider' and 'Gemma' in the article a genuine discrepancy or just unclear writing? - Why isn't the decision model offered via Bedrock instead of requiring more expensive SageMaker/Lambda infrastructure? **Highlights** > This is just the AI grift and uneducated hypebros. There is zero reason to use an LLM as a classifier. But AIbros with zero technical knowledge will push for it anyways even if it is 1000x slower and 1000x more expensive to run. > — [floro on hackernews · 5 comments](https://news.ycombinator.com/item?id=50436954) > I think this kind of change will pop up more often. In software engineering, we learned a long time ago that giant monoliths are not the answer. You want smaller, but specialized units that work together. They perform better at their specific task, are more maintainable, you can just upgrade one part of the system without having to impact the others, .... I don't know where the focus on building a single model that can do everything comes from. > — [autuni on hackernews · 2 comments](https://news.ycombinator.com/item?id=50437187) > This article is so confusing. For example, they say they tried the Strands Decider model and dismissed it for Gemma, then they continue to call the model they used Strands Decider? Did I miss something? > — [Gasp0de on hackernews · 1 comments](https://news.ycombinator.com/item?id=50437514) > Ok, but why did they chose Sonnet in the first place? Couldn't they have chosen Haiku or GPT Luna or maybe even some small cheap Chinese model? > — [psnehanshu on hackernews · 3 comments](https://news.ycombinator.com/item?id=50437307) > > JSON key order changed the answer. Sending the same content with keys in a different order moved the output probabilities by up to 0.19 Isn't it detrimental? > — [egorfine on hackernews](https://news.ycombinator.com/item?id=50437529) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50436640) · 219 points · 82 comments ## Similar posts on daily.dev - [AWS Open-Sourced a Decision Model as an Agent](https://daily.dev/posts/aws-open-sourced-a-decision-model-as-an-agent-wdbnayxnz) · Medium · 1 upvotes · 0 comments - [AWS Introduces Strands Decider 2B and Strands Box for AI Agent Development](https://daily.dev/posts/aws-introduces-strands-decider-2b-and-strands-box-for-ai-agent-development-dayqsdc2p) · InfoQ · 0 upvotes · 0 comments - [Cut AI token usage by 96%? Here’s how AWS Strands Agents does it.](https://daily.dev/posts/cut-ai-token-usage-by-96-here-s-how-aws-strands-agents-does-it--wt9gbklmo) · The New Stack · 1 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#aws-sagemaker](https://daily.dev/tags/aws-sagemaker) [View this post on daily.dev](https://daily.dev/posts/amazon-saves-millions-by-replacing-sonnet-4-6-with-strands-decider-bfmmo8l36) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Amazon saves millions by replacing Sonnet 4.6 with Strands Decider","url":"https://daily.dev/posts/amazon-saves-millions-by-replacing-sonnet-4-6-with-strands-decider-bfmmo8l36","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/amazon-saves-millions-by-replacing-sonnet-4-6-with-strands-decider-bfmmo8l36"},"datePublished":"2026-11-10T07:31:51.609Z","dateModified":"2026-11-10T09:33:06.363Z","description":"Amazon engineers replaced Claude Sonnet 4.6 with Strands Decider, a decision model, to label tickets in a large internal ticketing system, cutting costs by...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/44bc40885fd3b919c2feabd8cd79fa0e?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/44bc40885fd3b919c2feabd8cd79fa0e?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/amazon-saves-millions-by-replacing-sonnet-4-6-with-strands-decider-bfmmo8l36","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,llm,aws-sagemaker","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Amazon saves millions by replacing Sonnet 4.6 with Strands Decider"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/amazon-saves-millions-by-replacing-sonnet-4-6-with-strands-decider-bfmmo8l36#faq","mainEntity":[{"@type":"Question","name":"How much cheaper is Strands Decider than Claude Sonnet 4.6 for ticket classification?","acceptedAnswer":{"@type":"Answer","text":"Strands Decider E4B costs about $500 per million tickets versus about $22,200 for Claude Sonnet 4.6, a 98% reduction, roughly 44 times cheaper. Sonnet's cost comes from processing 6,724 tokens per ticket at $3 per million input and $15 per million output tokens, while Strands Decider uses about 2,010 tokens per ticket on a SageMaker ml.g5.xlarge instance at $1.408 per instance-hour. Comparing LLM versus decision-model costs for a classification pipeline is easier with real benchmarks tracked on daily.dev."}},{"@type":"Question","name":"What is a decision model and how does it differ from an LLM for classification tasks?","acceptedAnswer":{"@type":"Answer","text":"A decision model can't generate text; given a state and a question with a set of options, it returns a probability for each option plus a confidence score, rather than generating JSON or free text. This fits closed-choice tasks like ticket labeling well, where the state is the ticket and the options are a fixed label catalog, cutting both latency and token usage versus an LLM. Developers weighing decision models against LLMs for classification can follow comparisons like this on daily.dev."}},{"@type":"Question","name":"Does the order of JSON keys in a prompt affect a decision model's output?","acceptedAnswer":{"@type":"Answer","text":"Yes, sending identical content with JSON keys in a different order moved Strands Decider's output probabilities by up to 0.19. The fix was to always send requests with sorted keys. Anyone serving a decision model behind an API should check for this sensitivity before comparing runs or benchmarking results. Catching subtle serving quirks like key-order sensitivity matters when daily.dev readers deploy decision models in production."}}]} ```