Microsoft-Decision-1 is a decision model from Microsoft, available in Microsoft Foundry since 9 October 2026. It belongs to the category of System One models that we explained in our post Jev by TypeSafe: the AI model that doesn’t write text: instead of formulating an answer, the model scores predefined answer options and returns a calibrated probability for each. That makes it suited to the many small decisions in AI workflows, such as which model should handle a request, whether an agent action is permitted or how a ticket should be classified.
This page is based on the announcement on Microsoft’s Command Line blog by Achint Srivastava, VP of Software Engineering in the Office of the CTO, and the model overview on Microsoft Learn. This overview reflects the state as of October 2026.
What is Microsoft-Decision-1?
Per Microsoft, Decision-1 is a post-trained Qwen3.5-9B, optimised for fast, single-pass scoring. Microsoft plans to rebase the model on other foundation models in future, including models from Microsoft AI (MAI) and OpenAI.
The model answers structured questions:
- Yes/no: for example “May this agent perform the action?”
- Multiple choice and classification: for example assigning a ticket to one of several categories
- Ratings: scoring on a predefined scale
- Rubric-based grading: evaluating AI responses or agent actions against defined criteria
For each answer option, Decision-1 returns a calibrated probability. Applications can use it to set thresholds, for example acting automatically only at high confidence and passing uncertain cases to a human or a larger model.
How it works and limits
Per Microsoft Learn, Decision-1 processes inputs of up to 32K tokens and produces the decision scores in a single model call. The model is text only and does not accept images, audio or video. It produces no rationales: if you need a traceable explanation, combine Decision-1 with a language model or document the decision logic in the workflow itself.
Benchmarks per Microsoft
Microsoft states that it tested Decision-1 in a comparison across 36 benchmarks with nearly 150,000 questions kept out of training, covering routing, ranking, long context, multilingual and out-of-distribution tasks, reasoning and safety. According to Microsoft, Decision-1 achieved the highest accuracy and was also the fastest: 4.5 times faster than Quyet-1.0-Large, which achieved the second-best accuracy, and with a P50 latency about 35 times lower than GPT-6 Sol.
On robustness, Microsoft reports that the decision changes on 1.3% of perturbations and not at all when option descriptions are paraphrased or options are reordered. The safety evaluation covered 5,250 requests across 11 benchmarks on harmful content, jailbreaks and prompt injection.
From internal use, Microsoft reports among other things on Xbox Research: when labeling 10,000 pieces of open-ended feedback, Decision-1 was 14 times faster and 200 times cheaper than GPT-6 Sol. All figures come from Microsoft; independent evaluations are not yet available.
Pricing
| Model | Input (USD per 1M tokens) | Output | Availability |
|---|---|---|---|
| Microsoft-Decision-1 | 0.042 | free | Microsoft Foundry, OpenRouter announced |
Prices per Microsoft, as of October 2026. Because the model only returns probabilities for predefined options, input tokens effectively determine the cost.
EU availability and GDPR
Decision-1 is available in the Microsoft Foundry model catalogue. The region overview for Foundry models sold directly by Microsoft does not yet list the model with deployment types and regions. Whether an EU data zone or a region such as Sweden Central or Germany West Central is offered is therefore not documented. There are no open weights, so running it on your own infrastructure is not possible.
For workflows with personal data, check the concrete deployment region in your own Foundry project before use. If you need EU residency and prefer self-hosting, Cloudflare Clef and GLiNER2.5-Decide are open alternatives.
Where it fits among decision models
Decision-1 adds an offering straight from the Azure ecosystem to the still young category of decision models. Compared with the other models in this category:
- Jev from TypeSafe shaped the category and runs as a hosted API.
- Clef from Cloudflare is open (Apache 2.0), multimodal and can be self-hosted.
- Laya, Julia-1 and GLiNER2.5-Decide cover specialised and compact variants.
- OpenAI offers a similar approach with the Decisions API based on GPT-6 Luna.
- Microsoft-Decision-1 is proprietary, very low-cost and directly usable in Microsoft Foundry.
Our recommendation
For companies that already run their AI workloads on Azure and in Microsoft Foundry, Decision-1 is an obvious building block for routing, classification and agent controls: a few cents per million input tokens, calibrated probabilities and, per Microsoft, much lower latency than large language models. Before production use, we recommend clarifying the deployment region and evaluating the model against your existing solution with your own test cases, including German-language ones. In CompanyGPT and the AI Gateway, such decision models can serve as an upstream routing and checking stage; we support selection, evaluation and integration.
Related decision models
The category of decision models (System One) also includes Cloudflare Clef, Laya, Jev, Julia-1 and GLiNER2.5-Decide. OpenAI is moving in the same direction with the Decisions API based on GPT-6 Luna; details are on the OpenAI GPT page. More Microsoft models are listed under Microsoft Phi and MAI.