Microsoft launched Microsoft-Decision-1 on October 9, 2026, a model that scores fixed answer options instead of generating text. OpenRouter lists its input price at $0.042 per million tokens, with output free.
At a glance
| Company | Microsoft |
|---|---|
| Launched | October 9, 2026 |
| Available | Microsoft Foundry and OpenRouter |
| Price | $0.042 per million input tokens, output free, per OpenRouter listing |
| Context | 32,768 tokens |
| Base model | Post-trained from Alibaba's Qwen3.5-9B |
| Claimed accuracy | 83.5% across a 36-benchmark comparison, vendor-reported |
The signal
- Microsoft launched Microsoft-Decision-1 on October 9, 2026, a model that scores fixed answer options and returns a probability for each instead of generating text.
- The model is aimed at routing, classification, prioritization, verification and workflow control, where software can act on structured output without parsing chat replies.
- OpenRouter lists the input price at $0.042 per million tokens with output free, but buyers should test accuracy on their own labeled examples before automating decisions.
- Microsoft's 83.5% accuracy and speed claims are vendor-reported, and the base model is Alibaba's Qwen3.5-9B, which Microsoft says it will later move to its own and OpenAI models.
What happened
Microsoft's launch post introduces Microsoft-Decision-1 as a model for routing, classification, prioritization, verification and workflow control. It reads supplied content and a set of predefined answer options, then returns a probability for each option. The company says the model is built for these decisions rather than open-ended text.
The model was post-trained from Alibaba's Qwen3.5-9B. Microsoft has said it will later rebase the model onto its own MAI models and OpenAI models. Microsoft Foundry documentation lists the model for classification, routing, ranking, grading and binary decisions, and says it does not generate free-form responses or written rationales.
Microsoft reports 83.5% accuracy across 36 benchmarks, a figure relayed by and, and says the model ran 2.5 times faster than H2O-Lightning-4B v1.1 and 2.8 times faster than Jev. These are vendor-reported figures. The coverage reviewed does not include an independent reproduction.
By the numbers
Timeline
TypeSafe AI announced Jev, an early decision model that returns typed probabilities.
Microsoft launched Microsoft-Decision-1 in Microsoft Foundry and said it would be available on OpenRouter.
“Unlike LLMs, which are designed to generate text or reason through complex problems, decision models are purpose-built to deliver structured outputs that software can immediately act on.”
Who is affected
| Who | What changes | Impact |
|---|---|---|
| Engineering teams building agent pipelines | Can gate tool calls and check agent output with a probability instead of parsing chat text. | High |
| Support and operations teams | Can route tickets and rank incidents by severity at low cost, after testing accuracy on their own labeled examples. | Medium |
| Software buyers evaluating vendor claims | Must treat the 83.5% accuracy and speed multipliers as vendor-reported, since no independent reproduction appears in the coverage reviewed. | Medium |
| Teams already using Qwen-based or other open-weight models | Should note that Microsoft plans to rebase the model onto MAI and OpenAI models, which may change its behavior. | Low |
Why it matters
For software buyers, the main change is output shape. A decision model returns a fixed set of labels with probabilities, so an application can threshold or rank the result without parsing chat text. That fits routing tickets, filtering content, or checking an agent's output before the next step runs.
The price is the other draw. OpenRouter lists $0.042 per million input tokens with no output charge. Buyers should still test accuracy on their own labeled examples before automating a decision.
How a decision request works
- 1Supply state and questions
The application sends a state value, either text or JSON, plus one or more typed questions.
- 2Choose question type
Noul, choice or score sets the output shape: a yes/no probability, one label from a set, or a position on a scale.
- 3Single-pass scoring
The model reads the input and returns a probability for each option in one forward pass, with no generated text.
- 4Act on the result
The application thresholds or ranks the probabilities, and can escalate low-confidence results to a person or another model.
What to do now
For marketers
- Separate routing and classification tasks from copywriting tasks in your marketing workflows.
- Define fixed labels and clear criteria for a content classification pilot.
- Label sample content by hand for comparison with pilot results.
- Agree with IT on which uncertain decisions should go to a person for review.
For software buyers and IT teams
- Map a ticket-routing or output-verification workflow to fixed answer options.
- Test probability thresholds against your own labeled examples before enabling automated actions.
- Check decision API compatibility, deployment permissions and data-location requirements.
- Measure accuracy, latency and token costs against your current workflow.
Printable, one page, 19 checks. Enter your email and it opens right away.
We email you a copy. We never share your email and never send you sales emails.
Pick one. It takes 5 seconds and helps more people find these briefings. Your checklist opens straight after.
Have a work email? Use it instead and the checklist opens straight away.
Open it, then save it as a PDF or print it.
Know someone who would find this useful?
These steps are our recommendations based on this news, not statements from the companies involved.
What to watch
- Whether independent testers reproduce Microsoft's 36-benchmark accuracy and speed figures.
- Whether Microsoft moves the model's base from Qwen3.5-9B to its own or OpenAI models, as the company has said it will.
- The confirmed price on Microsoft Foundry, since the coverage reviewed cites only OpenRouter's listing.
Questions readers ask
What is Microsoft-Decision-1?
Microsoft-Decision-1 is a decision-scoring model that evaluates predefined answer options and returns a probability for each instead of generating free-form text. It is built for routing, classification, prioritization, verification, and automated workflow control.
How does Microsoft-Decision-1 differ from traditional LLMs?
Unlike generative models that write text, Microsoft-Decision-1 evaluates fixed options in a single pass and returns structured probabilities. This structure allows software to act on categorical choices without parsing natural language responses.
How much does Microsoft-Decision-1 cost?
Microsoft-Decision-1 is priced at $0.042 per million input tokens on OpenRouter, with free output tokens. The coverage reviewed does not include a confirmed price in Microsoft Foundry documentation.
What question types does Microsoft-Decision-1 support?
Microsoft-Decision-1 supports three question types: noul for binary conditions, choice for categorical selection, and score for ordered scales. Applications submit these typed questions alongside a state value formatted as text or JSON.
What base model is Microsoft-Decision-1 built on?
Microsoft built Decision-1 by post-training Alibaba's open-weight Qwen3.5-9B model. Microsoft stated it plans to later rebase the decision system onto its own MAI models and OpenAI models.
Where is Microsoft-Decision-1 available?
Microsoft-Decision-1 is available through hosted APIs on Microsoft Foundry and OpenRouter. The coverage reviewed indicates weights are not being distributed for local download, based on secondhand reporting.
Companies in this story
Sources
- ShatteredOriginal
- Command LineOriginal
- learn.microsoft.comOriginal
- openrouter.aiOriginal
- Pew Research CenterOriginal
- convovo.aiOriginal
- Liquid AIOriginal
- TechCrunch
- TechCrunch
- theregister
- theregister
Researched, fact-checked and edited by the Market Signal desk. Read our editorial standards. Spotted an error? Email hello@webvisiblity.com and we will correct it.