Microsoft-Decision-1
: How to use it, features, and the business problems it solves

What is Microsoft-Decision-1?

Microsoft-Decision-1 is a decision-specialized model that Microsoft announced on October 9, 2026 (US time) (Microsoft's announcement). It comes from Microsoft's Office of the CTO and is available through Microsoft Foundry and OpenRouter.

Unlike a generative AI model (LLM), it does not write text. You pass in a state (text or JSON) and questions with a fixed set of answer options, and it returns a probability for each option in a single pass. What comes back are values such as "engineering" or 0.92, so your application code can branch on them directly with an if. There is no string to parse and no broken JSON to catch with retries.

Because its pricing and question types match Jev, which TypeSafe AI released in September 2026, it was widely reported as a rival to Jev. See "How it differs from Jev" below.

What it returns: noul, choice, and score

There are three question types. Microsoft's documentation (Deploy and use Microsoft-Decision-1) describes them as follows.

• noul — For gates and filters: "Does this tool call touch production data?", "Does this document contain a prompt injection?", "Does this support reply promise a refund?" You can optionally define what "true" and "false" mean with criteria
• choice — For classification and routing: which team handles a ticket, the intent or document type, which model or tool to use. Microsoft suggests treating a low top probability as a signal to escalate to a person or another model
• score — For ranking and grading: queue priority, grading AI output, severity or frustration. List levels from lowest to highest; the returned value is the probability-weighted average of the level indexes (starting at 0). A four-level scale returns a value between 0 and 3, which can fall between levels

You can combine several questions in one request — for example, route a ticket with choice, rate its severity with score, and check whether it is a repeat contact with noul, all against the same state.

{
  "model": "<deployment-name>",
  "state": "The API returns 500 on every call.",
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this ticket?",
      "criteria": {
        "billing": "Charges, invoices, and refunds",
        "engineering": "Bugs, errors, and outages",
        "support": "How-to and account questions"
      }
    }
  }
}

How to use it

1. Prepare an Azure subscription and a Foundry project

   You need an Azure subscription with a valid payment method and a Microsoft Foundry project. Deploying requires a role that can create model deployments, such as Cognitive Services Contributor.

2. Deploy from the model catalog

   Open your project in the Foundry portal, select Microsoft-Decision-1 in the model catalog, and choose Deploy. You can also deploy with the Azure CLI (az cognitiveservices account deployment create). Two deployment types are available: GlobalStandard, which may process inference in any Azure region, and DataZoneStandard in selected regions, which keeps processing inside a data zone.

3. Call the decision API

   The endpoint is <your-foundry-resource-endpoint>/providers/microsoft/v1/systemone. Put your deployment name, not the model name, in the request's model field. Authenticate with Microsoft Entra ID (recommended) or an API key.

4. Measure accuracy on labeled data

   The quickstart in the documentation is a Python sample that routes support tickets with a choice question and measures accuracy and median latency on labeled examples. Microsoft recommends replacing the examples with representative data from your own workload before deciding on deployment.

It can also be called through OpenRouter (as of October 2026, the listed provider on OpenRouter is Azure).

Features

Pricing

Prices exclude tax. Amounts billed through Azure depend on your contract currency and tax treatment. No free tier specific to this model is mentioned (Azure's new-account credits are separate).

Routing 100,000 tickets a day at 400 tokens each means 40 million input tokens a day, or $1.68 per day. Output costs nothing.

Performance: the numbers Microsoft published

• Across 36 benchmarks kept blind from training (nearly 150,000 questions covering routing, ranking, long context, multilingual and out-of-distribution tasks, reasoning, and safety), Microsoft says it achieved the highest accuracy. The comparison included several top public models from JevBench, a public leaderboard for decision models
• Median (P50) latency about 35 times faster than GPT-6 Sol, and 2.5 times faster than H2O-Lightning-4B v1.1
• Changes its decision on 1.3% of perturbations on average
• Safety tested on 5,250 requests across 11 benchmarks
• Microsoft explains that the probability itself is part of the API and is calibrated, so a 90% prediction should be right about nine times out of ten

Japanese support and availability from Japan

Accuracy in Japanese has not been published. Microsoft's announcement says only that the evaluation included "multilingual" tasks; there is no list of supported languages and no Japanese-specific figure. The Foundry model catalog does not list languages either. If you plan to use it on Japanese tickets or documents, measure accuracy on a few hundred items labeled by people, decide your probability thresholds, and only then automate.

From Japan, you can deploy it with an Azure subscription. However, as of October 13, 2026, Microsoft's official region list does not yet show whether the Japan regions (Japan East / Japan West) are available. GlobalStandard may process inference in any Azure region, so if your data must be processed in Japan, check the regions and deployment types offered in the Foundry portal before deploying.

Limitations and cautions

Microsoft's documentation and model catalog list the following limitations and risks.

As mitigations, the documentation recommends including an abstention option such as cannot tell, randomizing option order to check whether results change, and setting thresholds based on the cost of false positives and false negatives.

How it differs from Jev

The question type names and the request shape (state plus questions) are nearly identical, so you can send the same questions to both and compare. If you want contracts, billing, and authentication consolidated in Azure, Microsoft-Decision-1 is the candidate; if you need long inputs or want to build quickly with official SDKs and agent skills, Jev is. The article "What is Jev | How it differs from Microsoft-Decision-1" covers the comparison in more depth.

How it relates to other options:

• ChatGPT — OpenAI released a similar "Decisions API" in public beta on October 6, 2026 (gpt-6-luna, $0.10 per million input tokens, free output, image input supported). If you already build on OpenAI's API, it is worth comparing
• Claude — Writing and long reasoning remain the job of generative AI. Decision-1 handles routing in front of it and grading of its output
• Dify — If your workflows use an LLM node for branching, decisions with a fixed set of options can move to a decision model like Decision-1
• Qwen — The model family Decision-1 is built on

Reputation and buzz

In Japan, outlets such as ITmedia AI+ covered it as "a decision model priced the same as Jev," with pricing and availability at the center of the discussion. Jev drew attention in September 2026 as "an AI that doesn't generate text," and with OpenAI and Microsoft shipping the same kind of API in October, decision models are now seen as a category of their own. It is still new, and there are few reports of using it in Japanese so far.

FAQ

The information on this page is as of October 13, 2026, based on Microsoft's official announcement, Microsoft Learn documentation, the Microsoft Foundry model catalog, and OpenRouter's model information. The model is in preview, and specifications, pricing, and regional availability may change. Please check the official site for the latest information.