Cut AI Labeling Costs 98% With Amazon’s Free Strands Decider
In short
- Amazon engineers replaced Claude Sonnet 4.6 with a small open model, Strands Decider, to label support tickets. They report 98% lower cost and 78% less time per call.
- Strands Decider is Apache 2.0, installs with pip and can run on a laptop CPU. It picks from options and tells you how sure it is. It cannot write text.
- The money angle: labeling, routing and yes/no checks do not need a big LLM. But below roughly 46,000 tickets a month a rented GPU costs more than the API.
Most of what small businesses use AI for is sorting. Is this email about billing or returns? Is this lead hot or cold? Is this review angry? You do not need a model that writes poetry for that. Amazon just showed what happens when you stop paying for one.
The old expensive way
You send every ticket to a big LLM with a long prompt: here are 20 categories, pick one. It works. It also costs tokens on every single message, and the answer takes seconds. Two Amazon engineers did exactly this with Claude Sonnet 4.6 on an internal ticketing system. Then they swapped in Strands Decider E4B and published the numbers on November 9.
What Strands Decider is
It is a small decision model from the Strands Labs group, released under the Apache 2.0 license. You give it a situation and a list of choices. It returns a probability for each choice and a confidence score. It cannot generate free text. That limit is the point: small, fast, cheap.
Here are the numbers from Amazon’s write-up:
| Per ticket | Sonnet 4.6 | Decider E4B |
|---|---|---|
| Time per call | 4,215 ms | 897 ms |
| Tokens | 6,724 | 2,010 |
| Estimated cost | $0.0222 | about $0.0005 |
| Per million tickets | about $22,200 | about $500 |
That is the 98% cost cut and the 78% time cut. Amazon says it adds up to millions of dollars of token costs a year. The setup was 20 labels plus a none option, an SQS queue, a Lambda function and a SageMaker endpoint on an ml.g5.xlarge instance.
Which size to pick
| Model | What you need | Speed (from the project page) |
|---|---|---|
| 2B (Qwen3.5 base) | RTX 3090, Apple-silicon Mac or CPU | 115 ms median on an RTX 3090 |
| E2B (Gemma 4) | about 9 GB of GPU memory, or CPU | not benchmarked reliably yet |
| E4B (Gemma 4) | about 15 GB of GPU memory in bf16 | 139 ms median on one H100 |
| 12B and 26B-A4B | larger GPU, CPU possible | see the project page |
Start with the 2B. It is the reference model and the easiest to run.
Set it up in 20 minutes
- Install. Run
pip install strands-decider. The E4B model needs the version from GitHub for now:pip install "strands-decider @ git+https://github.com/strands-labs/strands-decider". - Test one question. Run
strands-decider ask StrandsAgents/strands-decider-2B-qwen3.5-v1-2610 --state "Help! My payouts have been failing for 3 days!" --choice "Which team should handle this?=billing,sales,retail". The project page shows billing winning with 0.773. - Write your labels as short rules. One or two sentences each. Amazon had to shorten theirs because the model reads at most 4,096 tokens.
- Start a server. Run
strands-decider serve <model> --port 8000and send JSON requests to/v1/systemone. - Use the confidence. The project says answers at 0.9 or higher were right about 95% of the time on short classification tasks. Below 0.9, send the item to a person or to your big LLM.
- Run both models side by side first. Amazon used a three-way switch: old model, both, or new model. Compare for a week before you cut over.
- Sort your JSON keys. Amazon found that key order changed the output probabilities by up to 0.19. They now sort keys in every request.
Do the math: when does it pay off?
Be careful here, because this is where most people get it wrong. Amazon’s instance costs $1.408 per hour in us-east-1. Running it all month, about 730 hours, is roughly $1,028. Sonnet cost them $0.0222 per ticket. Divide: $1,028 / $0.0222 is about 46,000 tickets a month before the rented GPU wins. Below that, the API is cheaper.
A small shop has a better option: run the 2B on a machine you already own. Say 2,000 tickets a month. On an API at roughly €0.02 each that is about €40 a month, about €480 a year. On your own CPU or Mac, the model costs €0 plus a little electricity. Amazon’s per-ticket price is for their long prompts, so your real number will differ. Check it with your own prompts.
Where to use it
- Sort a shared inbox into returns, shipping, pricing and spam.
- Score inbound leads as call now, nurture or ignore.
- Flag angry reviews for a human reply.
- Route chat messages to the right person before a human reads them.
Run a one-week test before you trust it
Do not switch blind. Take 200 real messages from your inbox and sort them by hand into 4 to 6 boxes. That is your answer key. Then run the same 200 through the model and compare. How many match? How many of the misses were under 0.9 confidence, meaning they would have gone to a human anyway? If most errors sit there, the model is safe for your case. If errors show up at high confidence too, rewrite your labels as clearer rules or stay on your current model.
One easy trap: labels that sound alike. Returns and complaints often overlap. Merge them, or describe each with examples in two sentences. Most accuracy problems come from fuzzy labels, not from the model.
Limits and weak spots
- It outputs choices, yes/no answers and scores. No free-form text, no replies, no summaries.
- Confidence is calibrated only for short classification tasks.
- The Gemma models read text only. Vision needs the Qwen model with an extra install.
- The E4B install is from GitHub, not PyPI. Apple-silicon (MLX) support needs a clone for now.
- The project’s own notes say differences under about 10 tasks between single benchmark runs are unresolved, and the E2B score may be off because of a bug fixed later.
- Amazon’s saving is for one workload with long prompts. Yours may be smaller.
Alternatives
- A small LLM you already have. Use a cheap model with a short prompt and structured output. Easy, but you still pay per token.
- A classic classifier. With a few hundred labeled examples, a simple text classifier is free and fast. It needs your labeled data.
- Rules. If five keywords catch 80% of your tickets, start there and send the rest to a model.
Read also: Set Up Every AI Coding Agent From One File for €0 With Microsoft APM · Run a Team of AI Coworkers on Your Own Server With Rerun · DeepSeek Open-Sources an Agent That Answers Your GitHub Issues
Ivan’s Take
This is the kind of boring win I like. Stop paying a big model to pick one of five boxes. I would test the 2B on a real inbox this week, run it next to my current setup and only switch if the 0.9 confidence cases match. Do not rent a GPU for it unless you have tens of thousands of tickets.
Sources: Strands Agents · Strands Labs
HawkValue
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
