# What Is Strands Decider? A Small Decision Model Guide By Bartholomew Osei-Frimpong · 2026-11-10 · Source: https://www.activepieces.com/blog/what-is-strands-decider-a-small-decision-model-2026 --- Strands Decider is a small decision model from the Strands team (strands-labs on GitHub), announced by Amazon. Given some state and a question with options, it returns a probability for each option plus a confidence score, which workflow tools such as [Activepieces](https://www.activepieces.com) could use to route a step, per [Strands](https://strandsagents.com/blog/amazon-saves-millions-replacing-sonnet-with-strands-decider/). **It can't generate text**. Treat it as a classifier-style component, not a chatbot replacement. Keep in mind the following about this component: - It's new, and the field has had no time to prove it. - I'd vet it the way I'd vet any fresh component before wiring it into an automation: the maker's claims are a starting point, not evidence. - The maker describes the 2B version as built for fast experimentation and local development on its own site. - That wording is a positioning statement, not a reliability guarantee. ## Strands Decider Defined and Who Made It ### A decision in, probabilities and confidence out Picture a support ticket. The state you hand in is a short description: a customer reports a duplicate charge, the account is long-standing, and there are no earlier refund requests. The options you supply are refund, escalate to a human, or ask for more information. Back come one probability for each option and a single confidence score. In this invented illustration, refund would sit well ahead, escalation second and the request for more information last. That is the shape of the output, not real output from the model. The repository describes the mechanism as one forward pass with a slot or pointer head. In plain words, the model reads the state once and then points at one of the options you listed, rather than composing a sentence. That is why it is quick, and why it cannot answer anything outside your list. The confidence score is one number for how sure the model is about its pick. You can use it to route work: below a cutoff you choose, the case goes to a person. A high score doesn't prove the pick is right, and the pages as I read them don't promise the score is well calibrated. Your own labelled sample is what tests that. ### Strands Decider 2B model size and hosting needs The "2B" label is part of the model's name. Parameter count is a rough measure of size: more parameters need more memory and more compute for every answer. Two billion is small beside the large models it is meant to replace. That is why a model this size could plausibly run on modest hardware you control, instead of through a paid per-token service. ![Small box on a desk casting a huge machine's shadow](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/dcdebda3-b8a6-4737-9ad9-6486cfdcc3d1/what-is-strands-decider-a-small-decision-model-2-6a311010.webp) That is the whole cost argument, and it is a claim to price, not a fact to accept. The maker's pages do give figures: its E4B build used about 19 GiB of model memory on a SageMaker ml.g5.xlarge instance (one A10G, 24 GB) and took about 640 ms per request, so check those against your own hardware. ### Strands Decider sources and verified facts If an item is not on the list below, it is unverified here. The list holds everything I accept without testing, each item tied to a maker page. The facts are dated 2026-11-10, so a later change on those pages won't show up here. - Maker: the Strands team, under strands-labs on GitHub, per the [maker's site](https://strandsagents.com). - Announcement location: Amazon, per the [announcement post](https://strandsagents.com/blog/amazon-saves-millions-replacing-sonnet-with-strands-decider/). - Release date: the one shown on the [maker's blog](https://strandsagents.com/blog); I use no other. - Checked: 2026-11-10. ## What Strands Decider Does Inside an Agent Inside an agent run, the maker presents Strands Decider as a small model that makes the choice points. Those are choices such as which step or tool comes next, so the large model isn't paid to make them. ### What Strands Decider is built to do On the maker's pages it is a decision component, not a general assistant. That scope matters because you judge a decider on whether its choices are right and traceable, not on fluent text. ### The Amazon example as the maker tells it Treat the Amazon example as the maker's own illustration; the pages are the only source for it here. Read it as a claim about one scenario, chosen by the party selling the model, with no independent reproduction behind it. ![Chef presenting a dish judged only by an identical twin](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/9af30a3f-fd8a-4840-ab04-e00c38d6a56b/what-is-strands-decider-a-small-decision-model-2-9cedcd40.webp) Check which decision it covers, what the large-model baseline was, and whether the comparison is like for like. If any of those is unstated, the example shows what the model is meant to do, not how well it does it. On training the pages are silent, though they do report hardware for the E4B build: SageMaker ml.g5.xlarge with about 19 GiB of model memory. Parameter count alone doesn't give you a hosting bill, so the price comparison has to rest on what the maker publishes. Nor can you judge whether its training resembles your decisions, so only your own test cases can answer that. Neither gap is a defect. Both belong on your verification list. ## Strands Decider availability and cost You can get it from the maker's repository, and no price is published. You still pay for hosting, integration, testing and oversight, so price it against the hosted models it replaces. Get it from the maker's own repository and nowhere else. Anything you read about it elsewhere is secondhand, so confirm the release you install is the one the maker's page describes. Take the licence terms and hardware requirements from the repository's licence file and install instructions, not from a summary. Public code is not the same as an open source licence, and the maker's pages as I read them don't state one. Read the licence before you build on it. Terms on commercial use decide whether a pilot can become a product. ### Strands Decider pricing and unpublished costs With no price stated by the maker, the comparison point is the hosted models. [LLMCost](https://llmcost.tech/pricing-data) listed the older Claude Opus 4.6 at $5 to $25 per million tokens and GPT-5.5 at $5 to $30, which means neither model offered a clear price advantage. ![What the large hosted models Decider replaces charge](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/eaa680cf-8ac4-46df-bf6b-42845b484ba9/what-is-strands-decider-a-small-decision-model-2-254616d8.svg "Source: LLMCost") That spread means output-heavy workloads sit at the expensive end. Further down the list, GPT-6.1 Sol runs $2 to $10 and GPT-6.1 Terra $1 to $5. A small model has to beat Terra's floor to justify the migration, not just Opus's ceiling. **Free to download isn't free to run**. These are the lines you bear regardless: - hosting or compute for the model - integration engineering time - evaluation time on your own decisions - monitoring and human fallback review Weigh each against the token savings. Evaluation and fallback review scale with decision volume and risk, not with the model's size, so they can outweigh the per-token gap. ## Strands Decider strengths and limitations Before trusting the savings case, know that the evidence is thin and mostly the maker's own. It rests on claims you haven't yet checked, and review costs of that kind are what the per-token gap tends to hide. ### Strands Decider benchmark results from the maker In three difficulty tiers on its own benchmark, JevBench, Strands Labs scores the 2B model on its [evaluation page](https://github.com/strands-labs/strands-decider/blob/main/evaluation/README.md). JevBench is the maker's own test set, and its three tiers are easy, standard and hard. I haven't verified how tasks are sorted into them, so read the evaluation page before mapping the tiers onto your own decisions. The pattern matters more than any single score: the model is strongest where mistakes are cheapest, and weakest where they cost most. These are maker claims, not measurements anyone else has reproduced. Four gaps stand out: - The release is very recent, so there's little production history to learn from. - The one case study is a single Amazon deployment, and the maker reports it. - **No independent evaluation exists that I could find**. - The repository itself lists limitations, and you should read that list before relying on any headline number. Whatever one company saved, it isn't a result for yours. **Your mix of easy and hard decisions sets your accuracy, and the hard tier is where this model is weakest.** Ask what share of your decisions would land in the hard tier, since that share drives your error rate. Ask what a wrong decision costs and who catches it. Ask whether you can run your own labelled sample before switching. ## Strands Decider Compared With Established Options **Run your own labelled sample first**, because the table below can only tell you where Strands Decider sits against two familiar alternatives, not whether it gets your decisions right. In the table, Strands Decider is the only option whose maturity and cost basis we can't read from a source we trust. Take the empty cells as open questions to put to the maker, not as neutral entries. | | Maturity | Cost basis | Setup effort | Flexibility | Auditability | |---|---|---|---|---|---| | Strands Decider | Brand new | Not stated by the maker | Not stated by the maker | Not stated by the maker | Not stated by the maker | | Large general model (Claude Sonnet 5.5) | Established, current model line | Per-token usage | Prompt writing and a test set | High: new decision types need only a prompt change | Moderate: you can log inputs and outputs, but the reasoning is not a fixed rule | | Rules or classifier | Long-established practice | Engineering time, then cheap to run | Highest: rules written by hand, or training data labelled | Low: new cases mean new rules or retraining | High for rules, since each outcome traces to a rule; lower for a classifier | _Prices and plan limits checked against [strandsagents.com](https://strandsagents.com/blog/amazon-saves-millions-replacing-sonnet-with-strands-decider/) and [strandsagents.com](https://strandsagents.com/blog) and [github.com](https://github.com/strands-labs/strands-decider) on November 10, 2026._ When your decisions are narrow, repetitive, and checkable against labelled examples, a small model fits. When they shift often, or you must explain each outcome to an auditor, it fits poorly. If the maker won't state cost basis and audit behaviour, you can't show the saving over Sonnet. Nor can you show a reviewer why a case went the way it did. Ask for both in writing, then run the labelled sample. ## Connecting Strands Decider to a Workflow in Activepieces We haven't run the model in Activepieces, and it has no Strands Decider integration (checked 2026-11-10), so any connection is generic plumbing you build and test yourself. Three routes exist, simplest first: ![The Activepieces MCP Server configuration interface with setup instructions for Claude, Cursor, Windsurf and other clients.](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/ff7f3dff-f733-4e19-8254-96c00caccb0d/run-minicpm-v-4-7-locally-step-by-step-setup-gui-30d9d332.webp) 1. An HTTP request step, which needs a callable endpoint. 2. An AI step with a custom provider, which needs an OpenAI-compatible endpoint. 3. An MCP server, which needs a running server that exposes the model as a tool. For the HTTP request route, point a generic request step at whatever endpoint the maker documents. You write the payload and parse the reply, so nothing depends on Activepieces knowing the model, and since a run costs 1 credit however many steps it has, the extra request step doesn't raise your test bill. If the maker's pages show no endpoint, this route is closed. Only if Decider speaks the OpenAI chat format does the custom-provider route work, which you must confirm with the maker. If it does, the decision becomes an ordinary AI step you can swap against the large model, sitting in the same flow as your fixed rules, with one run trace for the whole execution rather than two systems bridged by a callback. ![A workflow builder showing a schedule trigger configured to run every Friday at 3 pm UTC.](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/6cfe18ef-b67b-43f6-b8bc-272af9e74826/elevenlabs-realtime-scribe-v2-pricing-and-featur-b82668c5.webp) ### Option 3: an MCP server This is the heaviest route, because you host and maintain a server. It makes sense only if an agent must choose when to call Decider, rather than a fixed step calling it every time. How large a test you can afford depends on cloud credits. [Activepieces](https://www.activepieces.com/blog/activepieces-cloud-pricing-plans-2026-price-guide) gives the Free plan ($0) 1,000 credits a month, so a small labelled sample fits only if each run costs few credits. Plus ($20) gives 10,000, enough to repeat that sample against both models. Team ($200) gives 50,000, which suits a larger set, so you can test against more models or scenarios. Check what each step consumes before choosing. ## Evaluating Strands Decider in one week Run your own past decisions through both models and let your error and cost numbers decide; one week is enough. The order below matters, because each step produces the input for the next. 1. Read the maker's pages. 2. Pull 50 to 100 past decisions with known outcomes. 3. Run them through your current model and Strands Decider. 4. Compare errors and cost per decision. 5. Keep a human fallback. 6. Recheck the pages. ### Day one: read the source pages Start with the [strands-labs repository](https://github.com/strands-labs/strands-decider). It describes a decider as one forward pass, with the answer read out through a slot or pointer head and "no streaming phase." So it picks from options you supply and doesn't write free text, which means any decision of yours that needs an open-ended answer is out of scope. Write down what the maker states and what it leaves unstated. Treat the gaps as open questions. Pull 50 to 100 past decisions where you know the correct outcome. Include the ugly edge cases, because a set of easy calls will flatter either model. Fix your tolerance for errors before you run anything, so the result can't bend your threshold. ![Tray of perfect cookies with a few lumpy ones added](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/e41d7602-a455-4217-bf4a-1fca193b1441/what-is-strands-decider-a-small-decision-model-2-a4c825af.webp) Compare error counts and cost per decision side by side. Keep a human fallback with a named owner for every case the model gets wrong or declines. Recheck the maker's pages after 2026-11-10, since a new model's claims can change under you. These results then set how much of the workload you hand over. ## Frequently asked questions Not everything a buyer needs before committing is on the maker's pages, and each gap changes how much you should trust the rest. The questions below stay open after the body, and every answer is limited to what the maker states. - Is it free? - Which licence does it use? Not stated, so you can't yet judge what you may modify or redistribute. - How big is it? The maker labels a 2B version, about 2 billion parameters, and reports about 19 GiB of model memory for its E4B build on a SageMaker ml.g5.xlarge instance at $1.408 per instance-hour, so you can price its hosting against the large model it's meant to replace. - Was it independently tested? Not stated, so every performance figure remains the maker's own claim. Four of the questions that come up most often follow, with answers kept to the same standard. Is Strands Decider open source? The maker's blog describes Strands Decider 2B as an open source decision model, though the exact licence terms are not stated on the pages, so read them before you build on it. Does Strands Decider only work with Strands Agents? Not stated as a restriction either way. Test it in the setup you actually run rather than assuming it carries over from the framework it's named after. Can Strands Decider replace a large language model entirely? No evidence supports that. It's positioned for decisions, and a model that makes the decision isn't one that drafts, summarises or converses. Does Activepieces support Strands Decider? Not as a built-in, and the maker's pages don't state support either, so any connection is an HTTP step you build and test yourself. ## Related reading - [d1 Ultra Decision Model: Specs, Pricing, and License](https://www.activepieces.com/blog/d1-ultra-decision-model-specs-pricing-and-license) - [Mistral Large 4: A Guide to the New Frontier Model](https://www.activepieces.com/blog/mistral-large-4-a-guide-to-the-new-frontier-model) - [Writer Palmyra API Pricing: Model Costs (2026 Guide)](https://www.activepieces.com/blog/writer-palmyra-api-pricing-model-costs-2026-guide) ## References - [Strands Labs](https://github.com/strands-labs/strands-decider/blob/main/evaluation/README.md) - [LLMCost](https://llmcost.tech/pricing-data) - [Activepieces](https://www.activepieces.com/blog/activepieces-cloud-pricing-plans-2026-price-guide)