Microsoft-Decision-1

Learning Objectives

• Explain what a decision model does and how it differs from a chat model
• Understand what Microsoft-Decision-1 returns and which kinds of question it can answer
• Judge Microsoft's benchmark claims, and know which comparisons it ran itself

What Is Microsoft-Decision-1?

Microsoft-Decision-1 is a small model that makes choices instead of writing answers. Microsoft released it on October 9, 2026, in a post by Achint Srivastava, a vice president of software engineering in the company's Office of the CTO, who called decision models "an important new category in AI."

The idea is the same one behind TypeSafe's Jev and OpenAI's Decisions API. A lot of AI work inside software is not a conversation. It is a small judgment: which team should get this ticket, whether this agent action is safe to run, how good this draft is. Asking a chat model for prose and then parsing a value out of it is slow and fragile. A decision model skips that step. You give it a fixed set of options, and it returns a probability for each one that your code can act on directly.

Microsoft built it by further training Qwen3.5, Alibaba's open model with 9 billion parameters, for fast scoring in a single pass. It says it will later rebuild the model on other foundations, including Microsoft's own MAI models and OpenAI's.

What It Answers

Microsoft-Decision-1 works through one structured API call. It supports four kinds of request:

• Yes or no — a probability that the answer is yes
• Multiple choice — a probability for every option you supply
• Rating — a score along a scale you define
• Rubric-based grading — judging an AI response or an agent's action against criteria you write

Because the options are fixed in advance, it cannot return something your code did not expect. That is a guarantee about the shape of the answer, not about whether the answer is right.

How Good Is It?

Microsoft says Decision-1 had the highest accuracy in its own comparison of 36 benchmarks, covering nearly 150,000 questions that were kept out of training. The rivals it named were Quyet-1.0-Large, H2O Lightning and GPT-6 Luna through OpenAI's Decisions API. It later updated the post to add Jev's results on accuracy and calibration.

On speed, Microsoft says it was the fastest model it measured: about two and a half times quicker than H2O Lightning, the runner-up, and about 35 times quicker than GPT-6 Sol at median latency. Speed matters more here than it sounds, because decisions often run in a chain. Microsoft's own example is that adding a tenth of a second to each of 20 decisions in a row adds two seconds to the whole task.

It also reports that across eight kinds of rewording of the same request, the model changed its answer 1.3 percent of the time, and never changed it when the options were reordered or their descriptions reworded. Safety testing covered 5,250 requests across 11 benchmarks for harmful content, jailbreaks and prompt injection.

Inside Microsoft, the company says:

• Xbox Research used it to label more than 10,000 pieces of player feedback, at quality it judged close to GPT-6 Sol while running 14 times faster and costing about 200 times less
• The Copilot team found it about 100 times faster than GPT-5.6 Luna
• A scientific-research team found its scoring 46 times more consistent than a chat-model approach, with replanning nearly four times faster

Availability

Microsoft-Decision-1 is available in Microsoft Foundry, Microsoft's platform for building with AI models. Microsoft's post says it is also available through OpenRouter, but it did not appear in OpenRouter's public model list as of October 11, 2026, so treat Foundry as the route that is confirmed. Microsoft has not published a context-window limit.

Pricing

The price is identical to TypeSafe's Jev: 4.2 cents per million input tokens, with output free. Since a decision model's output is a handful of probabilities, almost all of the cost is the text you send in.

Strengths

• Built for software, not conversation — fixed options in, one probability per option out, with no text to parse
• Very fast and cheap by Microsoft's measures, which matters most when many small decisions run one after another inside an agent
• Consistent answers: Microsoft reports almost no change in the decision when a request is reworded or its options are reordered
• Covers grading as well as routing, so one model can classify incoming work and also judge an agent's output against a rubric
• Sits inside Foundry, which suits organizations already buying Microsoft's AI stack

Limitations and Considerations

• The evidence is all vendor-reported. Microsoft picked the benchmarks and the rivals, and the Jev comparison arrived only after publication. Test it on your own decisions before trusting the headline numbers
• It only chooses among options you supply. It cannot write, summarize or reason through an open problem, so it complements a chat model rather than replacing one
• The base model is Alibaba's Qwen3.5. Some organizations restrict models built on Chinese-origin foundations, so check your policy, and expect behavior to shift when Microsoft rebuilds it on other models as promised
• Availability is narrower than announced. Foundry is confirmed; the OpenRouter listing was not visible as of October 11, 2026
• Same price as Jev, so price is not the differentiator. Choose between them on accuracy and calibration on your own data, and on which platform you already use

Best Use Cases

Getting Started

1. Open the model catalog in Microsoft Foundry and find Microsoft-Decision-1
2. Pick one decision your software already makes with a chat model, such as routing or a safety check
3. Write the options out explicitly, with a short description of each
4. Run a few hundred real examples through both your current approach and Decision-1, and compare accuracy as well as speed
5. Set a confidence threshold, and send anything below it to a person until you trust the model's calibration on your own data

Key Takeaways

• Microsoft-Decision-1, released October 9, 2026, returns a calibrated probability for each of a fixed set of options instead of writing text
• It is built on Alibaba's open Qwen3.5 model with 9 billion parameters, and Microsoft plans to rebuild it later on its own and OpenAI's models
• Microsoft says it led a 36-benchmark comparison and ran about 35 times faster than GPT-6 Sol, but every figure comes from Microsoft's own tests
• It costs 4.2 cents per million input tokens with output free, exactly what TypeSafe charges for Jev, so the two compete on accuracy and platform rather than price
• Foundry is the confirmed way to use it; the OpenRouter listing Microsoft announced was not visible as of October 11, 2026