Microsoft released Microsoft-Decision-1 on October 9, 2026 as a Foundry public preview, and Microsoft reports it ran 80 to 100 times faster than GPT-5 on one Xbox Research labelling job. It suits routing, classification and grading steps inside applications, so teams that need written answers or extended reasoning should keep a generative model.
Can you use it?
- Microsoft Foundry (Decision API): preview (Deploy from the Foundry model catalog; needs an Azure account that can be billed). Public preview.
- Free plan: not published.
Checked 31 Oct against Microsoft's own pages. Compare access across frontier models
What changed
- · Pricing change Microsoft-Decision-1: Microsoft Foundry (Decision API) plans: Deploy from the Foundry model catalog on an Azure subscription with a payment method → Deploy from the Foundry model catalog; needs an Azure account that can be billed; Price: $0 source
- · Release Microsoft-Decision-1 enters public preview in Microsoft Foundry (Microsoft-Decision-1 version 1) source
Microsoft-Decision-1 is a decision model in public preview on Microsoft Foundry, built on Qwen3.5-9B. It answers three typed question kinds (noul, choice and score) with probabilities instead of generated text, and Microsoft reports that Xbox Research sorted more than 10,000 feedback items with it at quality competitive with GPT-5.
Where it sits
- $0.032/M$ per 1M tokensBlended price (3:1)Lower is better#4 / 92peer median $1.70/Mvendor price, checked by HokAI
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Microsoft · Family: Microsoft Decision
Input modalities: text · Output: text
About Microsoft-Decision-1
Microsoft-Decision-1 is a decision model from Microsoft, announced on October 9, 2026 in the Microsoft Foundry blog and released in public preview in the Foundry Models catalog. Microsoft says it is built on Qwen3.5-9B, the open model line from Alibaba Cloud, and that future iterations will be rebased on OpenAI models and Microsoft's own MAI models. It belongs to a small group of models that skip free-form generation: Jev from TypeSafe AI uses the same request shape, and Intern-Decision and Cloudflare Clef target similar routing and judging work.
The model does not write answers. Microsoft's deploy guide says it makes a focused judgment about text or JSON and returns a typed, numerical decision, with no written rationale. A request carries a state (text or JSON) and one or more named questions. Three question types exist: noul returns a probability from 0 through 1 that a statement is true, choice returns the selected option plus a probability for every option, and score places an item on an ordered scale and returns a value with a probability for each level. The score is the probability-weighted average of the level indexes, so on a four-level scale it can land between levels, and Microsoft advises using scores for ordering and thresholds rather than as absolute, calibrated ratings.
Microsoft lists the intended jobs as agent controls, model routing, intent analysis, data labeling, AI judging, incident response routing, data validation and content classification. The one customer result in the announcement comes from Xbox Research, which used the model to sort more than 10,000 pieces of open-ended feedback into researcher-defined themes and reported quality competitive with GPT-5 at 80 to 100 times the speed. That is a vendor-reported figure from a single internal job, not an independent benchmark, and the post points to a separate Command Line blog for methodology. The post also says other Microsoft teams evaluated the model for judging Copilot response quality and for adaptive replanning in Microsoft Discovery.
Access runs through Microsoft Foundry. The deploy guide lists model version 1 with the DataZoneStandard deployment type in selected regions and GlobalStandard, under an API type called Decision. Calls go to a /providers/microsoft/v1/systemone route on your resource endpoint, authenticated with Microsoft Entra ID (recommended) or an API key, and the request's model value is your deployment name rather than the underlying model name. The announcement lists input at $0.042 per million tokens in both the US and EU Datazones, with output tokens shown as N/A. That input rate matches the one HokAI records for Jev, so a team choosing between the two is comparing behaviour on its own labels, not list price. The Jev alternatives guide covers the wider set of decision models.
Microsoft has not published a context window, latency figure or standard benchmark scores for this model in the two pages HokAI read, so those fields stay empty here rather than borrowing numbers from press coverage. The sensible test is the one Microsoft recommends: run representative inputs, compare decision accuracy on ambiguous cases, measure latency under your own traffic, price the cost per completed decision, and validate confidence calibration before wiring probabilities into an automatic gate or a human-review threshold. Teams that need written reasoning, long outputs or open-ended chat should keep a generative model for those steps and use this one for the high-volume choices around them.
Pricing
Input tokens are listed at $0.042 per million in both the US and EU Datazones in the Microsoft Foundry launch post of October 9, 2026; output tokens are listed as N/A. The model is a public preview, so Microsoft may change preview pricing. Billing runs through your Azure account, and the post says model requests incur applicable usage charges. Output tokens are listed as N/A, so no output rate applies in the published list. The same input rate is recorded for [Jev](/hub/models/jev), which also returns typed decisions.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0013 | $0 | $0.0013 |
| Support reply | $0.0001 | $0 | $0.0001 |
| One coding agent run | $0.0084 | $0 | $0.0084 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Three typed question kinds: A noul question returns a probability from 0 through 1, a choice question returns one option plus a probability for each option, and a score question returns a level on an ordered scale.
- State as text or JSON: A request carries a state that can be plain text or JSON plus named questions, so existing records can be passed in without a prompt template.
- Datazone and global deployment: The deploy guide lists DataZoneStandard in selected regions and GlobalStandard, with the launch post pricing the US and EU Datazones.
- Entra ID authentication: Sign-in goes through your tenant identity or a key issued on your own Foundry resource, so no separate vendor account is needed.
Pros
- Typed probabilities let an application threshold, rank or route without parsing generated text.
- Microsoft reports a single Xbox Research labelling job at quality competitive with GPT-5 and 80 to 100 times the speed.
- Only input tokens carry a listed price, and the US and EU Datazones share one rate, which keeps cost per decision simple to estimate.
Cons
- No written rationale, so explaining a decision to a user needs a second model.
- No published context window, latency or standard benchmark scores in the pages HokAI read.
- Public preview: terms, pricing and behaviour can change before general availability.
Frequently Asked Questions
How much does Microsoft-Decision-1 cost per million tokens?
The launch post of October 9, 2026 lists $0.042 per million input tokens for each of the two Datazones (US and EU), and shows output tokens as N/A. Microsoft says model requests incur usage charges on your Azure subscription. The input rate equals the one HokAI records for Jev, so the comparison between the two comes down to accuracy and latency on your own labels.
How do you deploy and call Microsoft-Decision-1?
Deploy it from the Foundry model catalog with the GlobalStandard deployment type, or DataZoneStandard where your region offers it, then send a state and typed questions to the Decision route under /providers/microsoft/v1/systemone on your endpoint. Sign in with Entra ID, which Microsoft recommends, or an API key; the model value in the request is your deployment name. Microsoft asks you to confirm the route and header against your own deployment before relying on its sample.
Is Microsoft-Decision-1 ready for production decisions?
Microsoft labels it a public preview, and the post tells teams to evaluate it on representative inputs, including ambiguous cases, before automating anything. Confidence estimates should be checked for calibration on your own data before they drive a review threshold. HokAI found no independent benchmark, so the Xbox Research result remains a single vendor-reported job.
How is Microsoft-Decision-1 different from a chat model such as GPT-5?
It does not generate free-form text or a written rationale. It returns a probability for a yes/no question, a probability per option for a choice, or a level on an ordered scale, which an application can threshold or route on. Microsoft frames it as the cheap, fast layer for repeatable decisions, with generative and reasoning models kept for drafting and analysis.
Who should use Microsoft-Decision-1, and who should pick something else?
It fits teams already on Azure that run high-volume routing, labelling, content classification or response grading and want a predictable response shape. Teams outside Azure, or those needing a published context window, rate limits or benchmark table, may find Jev or Intern-Decision easier to assess today. Anyone needing explanations in plain language should use a generative model.