Microsoft Decision-1: What It Is, Pricing, vs Jev and Laya

Microsoft-Decision-1 is Microsoft's decision model. Instead of writing text, it reads an input (a state, as text or JSON) and answers a set of typed questions with a probability for every allowed option: a yes/no gate, a pick from a fixed list, or a level on a scale. Microsoft announced it on October 9, 2026, on its Command Line blog. It is available in Microsoft Foundry for $0.042 per million input tokens, with output free. That is the same price as TypeSafe's Jev.

It belongs to the category TypeSafe started with Jev: System One models, which make fast, bounded decisions while a larger model does the slow reasoning. Microsoft-Decision-1 uses the same request body as Jev, but the weights are not published, so it runs only as a hosted Azure service. This page sums up what Microsoft has published, how to call it, and how it compares with Jev, OpenAI's Decisions API, Cloudflare Clef and the open-source Laya.

Microsoft-Decision-1 at a glance

Sources: Microsoft's launch post, the Foundry model catalog and the Microsoft Learn guide Deploy and use Microsoft-Decision-1. The Foundry catalog does not list the price; the $0.042 figure is from the launch post. Microsoft says it will later rebase the model on other base models, including Microsoft AI (MAI) and OpenAI models.

What Microsoft-Decision-1 is for

Microsoft describes it as a decision-scoring model for routing, classification, prioritization, verification and workflow control. The Learn guide maps the three question types to jobs:

• noul for gates and filters. Does this reply promise a refund? Does this document contain a prompt injection? Does this tool call touch production data? You get a probability from 0 to 1 and pick a threshold yourself.
• choice for classification and routing. Which team handles this ticket, which model or tool handles this request. You get the chosen option plus a probability for every option, so a low top probability can send the case to a person.
• score for ranking and grading. How severe is this incident, how good is this model answer. List the levels from lowest to highest. The result is the probability-weighted average of the level indexes, so on a four-level scale it runs from 0 to 3 and can fall between levels. Microsoft recommends using scores for ordering and thresholds, not as absolute ratings.

Several questions about the same state can go in one request, for example a team choice, a severity score and a repeat-contact noul.

Inside Microsoft, the launch post says the model sorted more than 10,000 open-ended feedback items for Xbox Research, checks Copilot answer quality and helps replan in Microsoft Discovery.

What Microsoft claims, and how to read it

From the launch post:

• Accuracy: the highest accuracy in a 36-benchmark comparison covering nearly 150,000 questions, none used in training. The comparison included Quyet-1.0-Large, Surogate Rune 26B-A4B, GPT-6 Luna Decisions, deck-31B, H2O-Lightning-4B and Strands-Decider 2B. An editor's note says accuracy and calibration results for Jev were added later, but the post's text gives no Jev numbers.
• Speed: median (p50) latency about 35 times lower than GPT-6 Sol and 2.5 times lower than the runner-up, H2O-Lightning-4B v1.1.
• Robustness: across eight kinds of input changes, the decision changed 1.3% of the time on average, and never when option descriptions were paraphrased or options were reordered.
• Internal results: on Xbox feedback, quality "competitive" with GPT-6 Sol at more than 14 times the speed and 200 times lower cost; on Copilot quality checks, "competitive" with GPT-5.6 Luna and 100 times faster; in Microsoft Discovery, 46 times more consistent than an LLM-based score.

How to read these numbers:

• They are all Microsoft's own. No independent benchmark had been published by October 13, 2026.
• The comparison leaves out close rivals. Cloudflare Clef and Perplexity's pplx-decider are not in the published list, Jev was added only in a later update, and the post's text gives no per-benchmark scores.
• Speed is compared with general LLMs. "35x faster than GPT-6 Sol" compares it with a model that writes text. Other decision models are built to be fast too.
• Microsoft lists its own limits. The Learn guide warns that scores can shift with how questions and options are worded or ordered, that calibration is strongest on familiar task types, and that the model should not be the only decision-maker for decisions about people.

How to call Microsoft-Decision-1

You need an Azure subscription with a Microsoft Foundry project. Deploy the model from the Foundry model catalog, or with the Azure CLI:

az cognitiveservices account deployment create \
  --name <ACCOUNT_NAME> \
  --resource-group <RESOURCE_GROUP> \
  --deployment-name <DEPLOYMENT_NAME> \
  --model-name "Microsoft-Decision-1" \
  --model-format Microsoft \
  --model-version "1" \
  --sku-name GlobalStandard \
  --sku-capacity 1

Then post a state and questions to your resource's decision endpoint. model is your deployment name, not the model name:

curl "$AZURE_ENDPOINT/providers/microsoft/v1/systemone" \
  -H "Content-Type: application/json" \
  -H "api-key: $AZURE_API_KEY" \
  -d '{
    "model": "'"$DEPLOYMENT_NAME"'",
    "state": "Checkout is down for all EU customers since 09:00.",
    "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"
        }
      },
      "severity": {
        "type": "score",
        "instructions": "How severe is this incident?",
        "criteria": ["Cosmetic", "Minor", "Major", "Critical"]
      }
    }
  }'

Microsoft recommends Microsoft Entra ID instead of an API key: send Authorization: Bearer <token> with a token for https://cognitiveservices.azure.com. The response has an answers object keyed by your question names, the underlying model name and token usage.

Moving from Jev

The request body is Jev's: a state, typed noul, choice and score questions, and criteria. Three things change: the URL (your Foundry resource plus /providers/microsoft/v1/systemone), the auth header (api-key or an Entra token), and model, which must be your deployment name. Code that builds the Jev body itself only needs those changes. If you use a Jev SDK, check that it lets you set the base URL and headers. Microsoft has not said whether it supports Jev's batch endpoint or its option and question limits, so test large requests first.

Microsoft-Decision-1 vs Jev, OpenAI, Clef and Laya

Accuracy claims in this category are self-reported and measured on different benchmark sets, so the table compares only published facts. See every model side by side in System One models.

Which one should you use?

• Your stack is already on Azure, with Entra ID, data zones and Azure billing: Microsoft-Decision-1 is the shortest path. Use DataZoneStandard if processing must stay in one region.
• You already call Jev and are happy with it: the price is the same, so switch only if Microsoft-Decision-1 is more accurate on your own labelled examples. Run both on the same few hundred cases first.
• You need images, video or audio in the decision: Cloudflare Clef, Clef-Flash or Clef-Omni. OpenAI's Decisions API takes images too. Microsoft-Decision-1 reads text only.
• You need open weights or must keep data on your own machines: Clef if you have a large GPU; Laya if you have a CPU, a laptop or no GPU at all. Laya can also be fine-tuned on your own labels, which closes much of the accuracy gap on a narrow task.
• You want the lowest hosted price: Clef-Flash on Workers AI ($0.038), then Jev and Microsoft-Decision-1 ($0.042).

FAQ

Is Microsoft-Decision-1 open source?

No. It is built on the open-weight Qwen3.5-9B, but Microsoft has not published its own weights. You can use it only as a hosted model in Microsoft Foundry (Microsoft also lists OpenRouter). If you need open weights with the same request format, look at Cloudflare Clef or Laya.

How much does Microsoft-Decision-1 cost?

$0.042 per million input tokens, with output tokens free, according to Microsoft's launch post. That is the same as Jev. The Foundry catalog points to Azure's model pricing page instead of listing a price, so check your region and deployment type there before you rely on it.

Is Microsoft-Decision-1 compatible with the Jev API?

The request and response bodies follow Jev's format: a state, typed noul, choice and score questions, and answers with probabilities. The endpoint URL, the authentication and the meaning of model (your deployment name) are Azure's, so you change those rather than only a model name.

Can I fine-tune or self-host Microsoft-Decision-1?

Not today. Microsoft's documentation covers deployment and calling only, with no fine-tuning option and no downloadable weights. To tune a decision model on your own labels, Laya has a one-command laya-train, and the Clef weights can be trained further on a GPU.

Should I use Microsoft-Decision-1 or Laya?

Use Microsoft-Decision-1 when you want a hosted model inside Azure with broad general knowledge and no infrastructure to run. Use Laya when you need decisions in milliseconds on ordinary hardware, want your data to stay on your own machines, or plan to fine-tune a small model for one narrow task.