Amazon Replaces Claude Sonnet 4.6 with Strands Decider, Cutting Ticket‑Labeling Costs by 98%
Bottom line
Amazon migrated a high‑volume internal ticket‑labeling system from Claude Sonnet 4.6 to the Strands Decider decision model, preserving labeling quality while cutting token consumption by 70% and per‑ticket cost from $0.0222 to roughly $0.0005 – a 98% reduction that translates to multi‑million‑dollar savings per year.
Why a decision model beats an LLM for ticket classification
A decision model does not generate free‑form text; it receives a fixed set of options and returns a probability distribution plus a confidence score. For Amazon’s ticketing workflow, the state is the ticket content and the options are the 20 predefined label categories. This matches the decision‑model interface perfectly, allowing the system to replace a generative LLM that produced JSON with a single‑token choice output.
"A decision model can’t generate text. It’s first given some state and a question with a set of options, then the decision model returns a probability for each option plus a confidence score." – Strands blog
Architecture of the production switch
The migration kept the existing AWS stack intact:
- Ticket ingestion – Events land in an SQS queue.
- Lambda trigger – A Lambda reads the ticket, queries a feature flag in AWS AppConfig, and calls either Sonnet, Strands Decider, or both.
- Model serving – Strands Decider runs on a SageMaker AI endpoint (ml.g5.xlarge) with the E4B Gemma‑4 model. The model occupies ~19 GiB RAM and processes a request in ~640 ms.
- Observability – CloudWatch logs, metrics, and alarms monitor latency and error rates; CDK defines the entire stack with Beta, Gamma, and Prod stages.
The three‑way feature flag (BEDROCK, DUAL, DECIDER) allowed Amazon to compare outputs side‑by‑side before fully cutting over.
Measured performance gains
| Metric | Claude Sonnet 4.6 | Strands Decider E4B | Relative improvement |
|---|---|---|---|
| Avg. latency per call | 4,215 ms | 897 ms | 78 % faster |
| Input tokens per ticket | 6,554 | 2,009 | – |
| Output tokens per ticket | 170 | 1 | – |
| Total tokens per ticket | 6,724 | 2,010 | 70 % fewer |
The token reduction stems from two factors:
- Strands Decider returns a single choice token instead of a JSON payload.
- The prompt is dramatically shorter; label descriptions were compressed into one‑ or two‑sentence criteria, fitting comfortably within the 4,096‑token window.
A day‑long Lambda‑duration trace showed average execution time dropping 74 % after the switch.
Cost comparison at published AWS rates
| Pricing component | Claude Sonnet 4.6 (Bedrock) | Strands Decider E4B (SageMaker) |
|---|---|---|
| Token price | $3.00 / M input, $15.00 / M output | $1.408 / instance‑hour |
| Per‑ticket cost | $0.0222 | ≈ $0.0005 |
| Per‑million tickets | ≈ $22,200 | ≈ $500 |
At ~640 ms per request, a single ml.g5.xlarge instance can handle ~5,600 tickets per hour, yielding a theoretical cost of $250 / M tickets at full utilization. In practice, Amazon observes about $500 / M tickets, still 44× cheaper than Sonnet.
Unexpected engineering detail
The team discovered that JSON key ordering altered the probability outputs by up to 0.19. They now canonicalize request objects by sorting keys before sending them to the decision model.
"One thing that surprised us was that 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. We now send every request with sorted keys."
When to consider a decision model
The switch succeeded because:
- The task is a closed‑choice classification over a static catalog.
- Ticket volume is high, making latency and token cost critical.
- The required context fits within a few thousand tokens.
If your labels need rich explanatory text, the catalog changes frequently, or the context exceeds the model’s window, a decision model may not be a drop‑in replacement.
Lessons learned from the rollout
- Three‑way feature flag: Running Sonnet and Decider in parallel (
DUALmode) gave real‑traffic disagreement data that offline tests missed. The flag also made rollback a 45‑second config change. - Cost‑driven refactoring: Re‑engineering pipelines to use cheaper, purpose‑built models can yield order‑of‑magnitude savings without sacrificing quality.
- Observability matters: End‑to‑end metrics (Lambda duration, SageMaker latency, CloudWatch alarms) were essential to validate the new path.
Community reactions
"About six years ago we built a ticket classification pipeline with SVMs that ran in a few milliseconds. It’s fascinating to see the field swing back to fast, cheap classifiers after a LLM detour." – armcat
"Saving ~$21,700 per million tickets implies >90 M tickets a year to hit $2 M savings, which feels plausible given the mix of escalations, alarms, and capacity requests." – Kaedon
"Decision models are a step up from LLMs for production workloads, but they’re still hard to work with and often require custom infra." – nlpnerd
"If you need a simple closed‑choice agent, try the Strands Decider repo and its serving notes on GitHub." – Strands blog authors
How to try it yourself
The Strands Decider codebase and inference guide are open‑source:
- Repository: https://github.com/strands-labs/strands-decider
- Inference documentation (including Gemma‑4 models): https://github.com/strands-labs/strands-decider/blob/0ac22a97e584f3ccf87cc9b8cc5fdbad57cdf534/docs/inference.md#gemma-4-models
Deploy the model on SageMaker, connect it to an SQS‑Lambda pipeline, and use AWS AppConfig for the three‑state switch to evaluate performance on your own traffic.
This post is based on the Strands blog entry “Amazon saves millions by replacing Sonnet 4.6 with Strands Decider” (published 2026‑11‑09) and the associated Hacker News discussion.
Sources
Related
- Project
- Dispatch
- Dispatch
- Dispatch
- Dispatch