Amazon Replaced Sonnet 4.6 with a Decision Model: 44x Cheaper, But Don't Rush to Copy It

Amazon replaced Sonnet 4.6 with a small decision model for 20-label ticket classification, cutting list cost ~44x and drawing the boundaries of specialized models.
On a 20-label ticket classification task, Amazon engineers replaced Claude Sonnet 4.6 with the small decision model Strands Decider, cutting list cost by roughly 44x. The point isn't "small model beats large model" but a precise boundary: decision models suit only narrow jobs with closed, enumerable outputs, stable label taxonomies, and high-frequency low-value calls, such as ticket routing, intent recognition, and content moderation grading; anything needing world-knowledge reasoning, natural-language explanations, or frequently changing labels still requires a general-purpose large model. The 44x figure is list cost — total cost must also account for engineering rework, monitoring, and retraining, and only high call volume turns the unit-price advantage into meaningful savings. The rollout pattern worth copying is single-task controlled validation, quantifying cost and quality before expanding scope; the core lesson is to select models by task complexity tier.
Large models aren't a panacea. Amazon engineers recently made a counterintuitive swap: on a 20-label ticket classification task, they replaced Claude Sonnet 4.6 with a small decision model called Strands Decider, cutting list cost by roughly 44x. The value of this case isn't the feel-good narrative of "small model beats large model" — it's that it draws a precise boundary: what kinds of tasks suit decision models, and what kinds can't live without large models.

What Task Was Actually Replaced
The task definition for this ticket classifier is extremely narrow: classify a user-submitted ticket into one of 20 preset labels. No open-ended Q&A, no multi-turn dialogue, no need to generate explanatory text — the output space is closed and enumerable.
This is exactly the home turf of a decision model. Models like Strands Decider don't pursue general language understanding; they treat classification decisions as a structured prediction problem — input text features, output label probabilities. With a small parameter count and short inference path, the compute cost per call is far lower than a general-purpose large model, and that's the main source of the 44x list cost gap.
The key distinction: Sonnet 4.6 was being used to do something it "can do but doesn't need to do." A general-purpose large model can certainly handle a 20-way classification, but the capabilities it mobilizes far exceed what the task requires. It's like using a freight truck to deliver a single letter — it gets there, but the per-kilometer cost structure is completely off.
What Strands Decider Is
Strands Decider is Amazon's internal small-model system for structured decision scenarios, designed with a different philosophy from general-purpose LLMs: it doesn't pursue open-ended language generation, but models the task as a discriminative problem — "given input features, select the optimal label from a finite candidate set." With a small parameter count and short inference path, each call performs a single forward-pass discrimination rather than autoregressive generation of a full text passage, so per-call compute cost and latency are significantly lower than a general-purpose large model. Such models typically require training or fine-tuning against a specific label taxonomy and have limited generalization, but on high-frequency tasks with fixed labels and clear boundaries, unit cost can be a fraction of a general model's. Understanding this is the only way to understand where the 44x cost gap comes from: the gap isn't that "the model is smarter," but that "the compute structure is cheaper."
The Capability Boundaries of Decision Models
The easiest way to misread this case is to take it as a signal that "small models will comprehensively replace large models." In reality, Strands Decider's applicability conditions are quite demanding.
What it's good at: classification tasks with fixed label taxonomies and clear boundaries; structured prediction with stable input/output formats; high-frequency calls that are latency- and cost-sensitive but don't require generation quality. Ticket routing, intent recognition, and content moderation grading all fall into this category.
What it's not good at: tasks requiring world-knowledge reasoning; scenarios requiring natural-language explanations; environments where the label taxonomy changes frequently and the model needs to generalize quickly; and any work that requires "understanding why" rather than "judging what." The original article explicitly emphasizes this is a "narrow job" — narrow-task fit, not a general replacement. In other words, what the decision model replaces is the large model's redundant capability on a specific pipeline, not the large model itself.
What Is a "Narrow Job"
A narrow job is a task whose input and output spaces are strictly bounded and whose decision criteria are enumerable: the input is relatively format-stable text, the output is one of a predefined finite label set, and the task doesn't require the model to explain its reasoning or generate new content. The opposite is an open task, such as Q&A, writing, or multi-turn dialogue, where the output space is nearly infinite and the model must mobilize world knowledge and language generation. The key feature of a narrow job is "closed verifiability" — every output can be exhaustively enumerated, monitored, and rule-validated, and error modes are enumerable and reversible. This is precisely why narrow jobs are suitable for migration to specialized small models; once a task needs to generalize to unseen labels or reason about "why," the premise of a narrow job breaks down and the decision model's advantage disappears with it.
The Rollout Pattern Worth Copying
What's more worth attention than the model choice is how this replacement was rolled out. The rollout pattern mentioned in the original points to a pragmatic path: first run a controlled validation on a single task, quantify the difference along both cost and quality dimensions, confirm the boundary, then expand the scope.
The rationale is that replacement risk on classification tasks is controllable. A closed output space of 20 labels means error modes are enumerable and monitorable, and problems in production can be quickly rolled back. If it were an open-ended generation task instead, the same replacement decision would require far more caution.
For teams currently doing model selection, this case offers an actionable checklist: Is the task output closed and enumerable? Is the label taxonomy stable? Are there a large number of high-frequency, low-value calls? When all three answers are yes, decision models are worth bringing into the evaluation scope.
The Cost Math Must Be Complete
44x is the gap in list cost, and this metric deserves attention. Actual total cost also includes engineering rework, building the monitoring system, and model retraining overhead when the label taxonomy changes. For scenarios with low call volume, these fixed costs can eat up the unit-price advantage.
What really makes decision models worthwhile is high-frequency call scenarios — hundreds of thousands of classification requests per day, where the unit-price gap is amplified by call volume into substantial absolute savings. In low-frequency scenarios, sticking with a general-purpose large model is actually the simpler choice.
Implications for Model Selection
The core conclusion of this case can be compressed into one sentence: select models by task complexity tier, rather than configuring uniformly around the upper bound of model capability.
A general-purpose large model's capability surplus is an advantage in the exploration phase and a cost in the production phase. Once task boundaries have been repeatedly validated and the output space has converged, migrating such work to specialized small models is trading redundant capability for real money. But the precondition for migration is that the task is narrow enough, stable enough, and high-frequency enough — if any one of the three conditions is missing, the cost-effectiveness of the replacement must be recalculated.
The significance of Amazon's replacement isn't proving that decision models are stronger than Sonnet 4.6, but demonstrating a kind of engineering discipline: let each model do only what it should do.
Related articles

AI Writes Your Code—Do You Still Need to Learn Git? Five Reasons
AI coding agents handle Git operations automatically—is it still necessary to learn Git? Five reasons: team communication, reviewing agent PRs, monitoring code changes, fallback when AI goes down, and the joy of learning—plus how the focus of learning Git is shifting in the agent era.

Anthropic Reportedly Prices IPO at $83 Per Share, Valuing Company at $1.5 Trillion
Anthropic reportedly prices IPO at $83 per share, above the expected range, with a fully diluted valuation of $1.5 trillion. This article analyzes the signal significance of the pricing and its impact on the AI industry's financing environment.

Open-Source Code Hosting Platform Codeberg Suffers Outage, Sparking Debate in Developer Community
Open-source code hosting platform Codeberg.org suffered an outage, sparking Hacker News debate over nonprofit platform stability and open-source infrastructure dependency risks.