Microsoft-Decision-1
Microsoft · Productivity · Business
Microsoft-Decision-1 is a decision scoring model in Microsoft Foundry that returns a calibrated probability for each fixed answer option instead of writing a response.
About Microsoft-Decision-1
What Is Microsoft-Decision-1
Microsoft-Decision-1 is a model for structured decisions. You send it a situation and a question with a limited set of answers. It sends back a number for every answer, and software can act on that number without parsing a paragraph of text.
The model sits in Microsoft Foundry and on OpenRouter. Microsoft built it on Alibaba's open-weight Qwen3.5-9B and post-trained it for single-pass scoring. That makes it small next to a frontier chat model. It is also cheap to run, which is why an AI decision model now competes with prompt calls to a chat LLM for routine routing work.
The work it replaces is the quiet kind that fills an agent's day. Which team should get this ticket. Is this reply promising a refund. How severe is this outage. Those calls normally go to a large language model, which spends a second writing an explanation nobody reads. A decision model skips the explanation.
The limit is scope. Microsoft-Decision-1 only chooses among options you define. Ask it to draft a policy or weigh evidence nobody has framed as a question, and it has nothing to score. It also returns no reasoning, so a wrong answer takes more digging to diagnose than a chat model's wrong answer.
Getting Started
- Open the Foundry portal, go to your project, and find Microsoft-Decision-1 in the model catalog.
- Select Deploy and pick a deployment type. DataZoneStandard keeps inference inside one data zone. GlobalStandard draws on capacity from any supported region.
- Copy the endpoint, the deployment name, and an API key, or connect with Microsoft Entra ID.
- Send a request with a
state(plain text or JSON) and one or more typed questions. Usenoulfor yes or no,choicefor one label from a fixed set, andscorefor an ordered scale. - Read the
answersobject, apply your own threshold, and log the probability next to the decision so you can audit it later.
Product Information
A quick look at Microsoft-Decision-1's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- Platform and backend engineers
- Agent builders
- Data and support operations teams
- AI quality leads
Tasks
- Ticket triage
- Model routing
- Response grading
- Safety gates
Scenarios
- An agent has to choose its next action from a short menu on every turn. Scoring four options costs a fraction of a chat completion and returns in one pass.
- A support team receives more tickets than reviewers can read
- A research pipeline screens candidate experiments
Key features
Calibrated probabilities instead of prose
Every answer comes back with a probability between 0 and 1, and Microsoft treats that number as part of the API rather than a ranking hint. A 90% prediction should be right about nine times out of ten on representative cases. That lets an application set a threshold, act above it, and hand the rest to a person.
Three question types for different decisions
The noul type answers a yes or no question with a single probability. choice picks one label from a fixed list and returns a score for every option. score places an item on an ordered scale and returns a probability-weighted average of the levels. You can send several questions in one request and get them all scored in the same pass.
One pass over up to 32K tokens
The model handles an input of up to 32,768 tokens and returns structured JSON in a single invocation. Long instructions, a ticket thread, or a chunk of source code fit in that window. There is no back-and-forth, so a decision costs one request rather than a conversation.
Decisions that survive rewording
Production inputs get paraphrased, reordered, and reformatted. Microsoft tested the model by perturbing the same request eight ways and measuring how often the answer flipped. The average was 1.3%, with no flips when option descriptions were paraphrased or options were reversed.
Low latency at $0.042 per million input tokens
Microsoft reports a P50 latency about 35 times faster than GPT-6 Sol, and 2.5 times faster than H2O-Lightning-4B v1.1, the runner-up in its comparison. Input tokens cost $0.042 per million and output tokens are free. Pricing that low changes what you can afford to score, such as every message in a busy queue.
Built for AI agent evaluation and control
Microsoft positions the model for AI agent evaluation: grade a proposed action against a rubric, check groundedness against supplied evidence, and let the agent continue, retry, or stop. It also supports explicit abstention, so the model can pick "cannot tell" when the evidence is too thin to decide.
Pros and cons
Pros
- Output is typed JSON with a probability per option, so no parsing layer or output validator is needed.
- The price of $0.042 per million input tokens makes high-volume AI classification practical, even if you score every message in a busy queue.
- Latency is low enough to sit inside a loop that runs several times per task.
- Answers hold up when inputs are reworded, which matters for prompts that users write in their own words.
- Safety testing covered 5,250 requests across 11 benchmarks, including jailbreak attempts and prompt injection.
Cons
- It returns no explanation, so a bad decision gives you a number and little else to debug.
- Questions must have a fixed set of options. Open-ended judgment calls fall outside what it can do.
- The model is in preview, so Microsoft offers no production SLA and the API may change.
- There is no free tier. You need an Azure subscription with a payment method before you can deploy it.
Frequently asked questions
It handles structured decisions inside software. Common jobs include request classification, ticket routing, queue ranking, response grading, and picking the next step for an agent. Each call returns a probability for every option you supply.
Related content
Explore related tools, skills, and articles for Microsoft-Decision-1.
Microsoft-Decision-1 Alternatives
Base Code
Base44 (Wix) · Coding · ProductivityBase Code is a shared cloud development environment from Base44. Teammates describe changes to an existing GitHub codebase in chat, preview them live and send them back as pull requests.
Knocket
Knocket (Tencent RTC team) · Productivity · Business · ChatbotKnocket is a free live chat and contact page tool that lets visitors message you from your website, app, or a shareable link, with every reply handled from Telegram or one inbox.
FloorGen
Vinit Kumar (FloorGen) · Productivity · China AIFloorGen is a browser-based floor plan maker that turns a sentence or a sketch into an editable 2D plan and a 3D walkthrough. Export to PDF, DXF or video, and draw without an account.