Microsoft has released Microsoft-Decision-1, an AI model that does not write text. It reads a piece of text or data, answers a fixed question about it and returns a probability score that other software can act on. Microsoft announced it in a post dated 9 October 2026 in the United States, which was already 10 October 2026 in Australia, and says it is available in Microsoft Foundry and through OpenRouter.
What happened
Microsoft's post introducing Microsoft-Decision-1, written by Achint Srivastava, VP of Software Engineering in the Office of the CTO, calls it the company's new model for fast decision-scoring. Microsoft says it is designed for routing, classification, prioritisation, verification and workflow control.
The post places it in a category Microsoft calls decision models. A large language model (LLM) is designed to generate text or reason through complex problems. A decision model, Microsoft says, is purpose-built to deliver structured outputs that software can immediately act on.
In practice, you give the model a fixed set of answer options and it returns a calibrated probability score for each one. Calibrated means the confidence figure is meant to match how often the model is right. In Microsoft's words, "a 90% prediction should be right about nine times out of 10 on representative cases".
The model supports yes/no, multiple-choice and rating questions, plus rubric-based grading of AI responses and agent actions, through a structured API call (an API is the way one program asks another to do something).
Microsoft says it built the model by post-training Qwen3.5-9B, and that it plans to rebase it on other models, including Microsoft AI (MAI) and OpenAI models. The post gives no date for that change. The Register's report of 10 October 2026 describes the Qwen family as developed by Alibaba Cloud, the cloud computing arm of Alibaba.
The Register also counts more than 100 such models, from vendors including OpenAI, Cloudflare, Snowflake and H2O.ai, and says Jev, the TypeSafe AI model its headline casts Microsoft as challenging, was announced three weeks earlier.
Key details
| Item | What Microsoft says |
|---|---|
| Model | Microsoft-Decision-1, a model for fast decision-scoring |
| Announced | Post dated 9 October 2026 (United States), 10 October 2026 in Australia |
| Where | Microsoft Foundry and OpenRouter; the post does not list countries, regions or plan names |
| Built on | Qwen3.5-9B, post-trained by Microsoft; a rebase on MAI and OpenAI models is planned |
| Price | US$0.042 per million input tokens; output tokens are free |
| Speed claim | 2.5 times quicker than H2O-Lightning-4B v1.1, the runner-up, and 35 times quicker than GPT-6 Sol |
| Accuracy claim | Highest accuracy in a 36-benchmark comparison spanning nearly 150,000 questions |
| Consistency claim | Changes its decision on 1.3% of perturbations on average, when the same request is altered in eight ways |
| Safety testing | 5,250 requests across 11 benchmarks covering harmful content, jailbreak attempts and prompt injection |
Every figure in the table is Microsoft's own. The post also describes an internal trial: XBOX Research used the model to sort more than 10,000 open-ended pieces of feedback and reviews into a fixed set of themes, and found it competitive on quality with GPT-6 Sol while running over 14 times faster and 200 times less expensive.
Why it matters
Many everyday business tasks are not writing tasks. They are sorting tasks: which team should handle this enquiry, whether a record meets the requirements, which label a piece of customer feedback belongs under. Microsoft's list of uses covers this ground: data labelling, intent analysis, data validation, and content classification and filtering.
A fixed answer with a number attached is something software can compare against a threshold, and speed and cost matter when a workflow has many such steps. Microsoft's post notes that adding 100 milliseconds to each of 20 sequential decisions adds two seconds to a workflow. Tokens are the units of text an AI service counts for billing, and here only the input is charged.
The other reason is control over AI agents, the tools that take actions on your behalf. One use Microsoft lists is agent controls: scoring an agent's proposed next step and deciding whether it should continue, stop, retry or hand off to a model, a tool or a human. We covered the wider question in our guide to permissions and data access for AI agents in a small business.
There are limits, and Microsoft's documentation for the model lists them plainly. The model does not explain itself: it returns a typed, numerical decision with no written rationale. It might reflect biases from its base model and training data. Scores can change with how you phrase or order the questions and options, and a poorly framed question still returns a score.
What this means for businesses
Microsoft-Decision-1 is a building block, not an app you sign in to. The documentation lists an Azure subscription with a valid payment method, access to Microsoft Foundry and a Foundry project among the prerequisites. Most small and medium businesses will meet a model like this as one step inside software that someone builds or connects for them.
It suits high-volume, repetitive choices with a fixed set of answers. If you run a wholesale business that receives hundreds of emails a week, sorting each one into orders, returns, accounts or supplier queries is that kind of task. It does not suit work that needs writing. Microsoft's documentation says to use a generative LLM when the application needs to create, summarise, rewrite or explain content, and that the two can be combined.
What to do or check:
- List the decisions first. Write down the choices your team repeats that have fixed answers. business.gov.au's guidance on AI says to start from the business problems, not the tool, and to trial AI in 1 or 2 areas.
- Test on your own examples. Vendor benchmarks are a starting point. Microsoft's documentation says to validate the model on data that is representative of your use case.
- Set a threshold and an escape route. The documentation suggests confidence thresholds based on the cost of false positives and false negatives, an option such as "cannot tell", and passing a decision with a low top probability to a person or another model.
- Keep people in decisions about people. For matters such as credit, employment, housing or legal questions, Microsoft says to use the model for decision support with meaningful human review, not as the sole decision-maker, and to tell affected users when AI contributes.
- Check what data goes in. The OAIC's guidance on commercially available AI products says privacy obligations apply to any personal information put into an AI system and to output that contains personal information.
- Name who is accountable. business.gov.au says your business is responsible for everything its AI tools do, and recommends making someone high up in the business accountable for AI safety.
Our business systems and integrations and custom software work covers this kind of routing and automation, and our IT consulting service can help you decide whether it is worth doing at all. We offer a free first consultation and a written quote before any work starts: send us an enquiry.
Frequently asked questions
What is Microsoft-Decision-1?
It is an AI model from Microsoft built for fast decision-scoring. Given a fixed set of answer options, it returns a probability score for each option instead of written text, so other software can act on the result.
What does Microsoft-Decision-1 cost?
Microsoft's post of 9 October 2026 gives the price as US$0.042 per million input tokens and says output tokens are free. The post does not list regions or plan names, so confirm the current price on Microsoft's own pages before you budget.
Can a small business use Microsoft-Decision-1 without a developer?
Not directly. It is called through an API, and Microsoft's documentation lists an Azure subscription with a valid payment method, access to Microsoft Foundry and a Foundry project as prerequisites. Most small businesses would meet it inside software built or connected for them.
Which model is Microsoft-Decision-1 built on?
Microsoft says it built the model by post-training Qwen3.5-9B. The Register describes the Qwen family as developed by Alibaba Cloud. Microsoft says it plans to rebase the model on others, including Microsoft AI (MAI) and OpenAI models, and gives no date.
Should a decision model make decisions about people?
Not on its own. Microsoft's documentation says that for consequential decisions about people, such as credit, employment, housing or legal matters, the model should support the decision with meaningful human review, and affected users should be told when AI contributes.
Key takeaways
- Microsoft-Decision-1 returns probability scores for fixed answer options. It does not write text or explain its reasoning.
- Microsoft announced it in a post dated 9 October 2026, with availability in Microsoft Foundry and through OpenRouter.
- The price in the post is US$0.042 per million input tokens, with output tokens free. The speed and accuracy figures are Microsoft's own.
- It fits repetitive sorting, routing and checking tasks. Writing, summarising and explaining still need a generative model.
- Test it on your own data, set confidence thresholds, and keep human review for decisions about people.