Most AI launches promise longer answers, but Microsoft Decision 1 promises fewer words and better choices. For example, it does not write a reply or explain its reasoning in prose. However, that restraint is the point, because many production systems do not need a paragraph. Therefore, this guide explains how a decision model works, what it costs, and where it earns a place in real automation.
In addition, Microsoft announced the model on October 9, 2026. Microsoft Foundry gained a new class of small, specialised models. As a result, developers can pair a large chat model with a cheap scorer that checks, routes, and gates each step.
What Is Microsoft Decision 1?
Microsoft Decision 1 is a decision scoring model, not a chat model. For example, you give it a situation, a question, and a fixed list of options, and it returns a calibrated probability for each option as JSON. In addition, it supports yes or no questions, multiple choice questions, ratings, and rubric based grading. Therefore, software can act on the result at once, without parsing free text.
The model also offers an abstention option, so it can say it cannot tell instead of forcing a confident guess. For example, a support router can send a low confidence ticket to a human, while a high confidence ticket goes straight to the right team. However, teams must set those thresholds themselves, because the model returns numbers, not business rules.
The Qwen Base Behind the Model
Microsoft built the first version by post training Alibaba Qwen3.5-9B for single pass decision scoring. In addition, Microsoft plans to rebase later versions on other models, including Microsoft AI and OpenAI models, but it has not set a date. Therefore, buyers should judge the current model on its current behaviour.
Microsoft describes Microsoft Decision 1 as a fast decision scoring model in the official Microsoft announcement, and it positions the model for routing, classification, prioritisation, verification, and workflow control. For example, the same scoring step can route a ticket, grade an agent action, or check whether an answer matches supplied evidence.
How Decision Scoring Works in Practice
Microsoft Decision 1 flips the usual chat workflow. For example, a developer defines the options before the model runs, such as billing, technical, or sales, and the model scores each option. However, the developer must write clear option descriptions, because vague options produce vague scores. Therefore, good option design matters more than prompt tricks.
In addition, the output format keeps downstream code simple. The model returns structured JSON with a probability for each option, so code can compare scores, apply a threshold, and trigger an action. As a result, engineers avoid fragile text parsing, and they can log every score for audits.
Decision Model vs LLM: Why the Difference Matters
A decision model vs LLM comparison starts with the job, not the size. For example, a large language model generates open ended text, while a decision model answers a closed question with numbers. However, a large language model can also answer a multiple choice question, so teams may ask why they need a separate scorer. Therefore, the answer comes down to cost, speed, and control.
In addition, decision models reduce a common failure mode. A chat model may change its wording or invent a category, and downstream code must handle every variation. As a result, a scorer that returns only probabilities makes automation more predictable. Meanwhile, a chat model still wins for drafting and explaining a complex case.
Key Features
Microsoft Decision 1 features focus on speed, structure, and robustness. For example, Microsoft reports a median latency of about 85 milliseconds in its own tests, and it claims about 35 times faster performance than GPT-6 Sol in the same harness. However, those figures come from Microsoft, so teams should measure latency on their own traffic first.
In addition, Microsoft reports strong robustness to option order and wording changes. For example, decisions flipped on only 1.3 percent of perturbed requests on average in Microsoft tests, with no flips when testers reversed or shuffled options. Meanwhile, teams should still add their own safety checks for sensitive workflows.
Microsoft Decision 1 Pricing Explained
Microsoft Decision 1 pricing is simple, and that simplicity helps budgeting. For example, Microsoft charges 0.042 dollars per million input tokens, and it charges nothing for output tokens. In addition, the output stays tiny by design, because the model returns scores, not prose. Therefore, teams can estimate cost from input volume alone.
To see the scale, consider a routing system that sends 1,000 tokens per decision. For example, one million such decisions would use one billion input tokens, and the input cost would total about 42 dollars at the listed rate. However, real systems also pay for the surrounding chat model and monitoring. As a result, the scorer can still save money overall, because it can replace expensive chat model calls for routine checks.
Comparison Table: Rivals at a Glance
| Model | Job | Input price per million tokens | Output price | Context | Best for |
|---|---|---|---|---|---|
| Microsoft Decision 1 | Decision scoring | 0.042 dollars | Free | 32,768 tokens | Microsoft Foundry automation |
| TypeSafe Jev | Decision scoring | 0.042 dollars | Free | 64,000 tokens | Decision scoring outside Azure |
| OpenAI Decisions API | Decision scoring | 0.10 dollars | None charged | Large | OpenAI centred teams |
| General large language model | Open text generation | Higher | Charged | Large | Drafting and open reasoning |
For another specialised model built for agents, see our guide to NVIDIA Nemotron 3.5, because it shows how a multimodal model can support agent workflows that need vision as well as text. In addition, if raw speed matters for voice or real time work, our Mercury 3 coverage shows how a diffusion model changes latency. Therefore, choose a decision scorer for a fixed choice, and a general model for open text.
Strengths and Limitations
Microsoft Decision 1 strengths start with price, speed, and a clean output contract. For example, teams can run thousands of routing, grading, and verification calls without paying chat model rates for long replies. In addition, the structured JSON output simplifies code and audits, and the abstention option gives teams a safe path for uncertain cases. Therefore, the model suits the hidden layer of automation, where reliability matters more than eloquence.
However, the limitations deserve equal attention. First, the model accepts text only, so it cannot score images, audio, or video directly. Second, Microsoft does not release the weights, so teams must use the hosted API and cannot self host the model. Third, the context window reaches only 32,768 tokens, so very long documents need chunking first. As a result, teams with strict self hosting rules should check whether a hosted API meets their policy.
In addition, all headline accuracy figures come from Microsoft, not from independent production tests. For example, Microsoft reports 83.5 percent average accuracy across 36 benchmarks covering nearly 150,000 questions, but real customer data always differs. Therefore, teams should build a labelled test set and set thresholds before they automate consequential gates such as refunds or account actions.
Who Should Use It
Business automation teams should test the model first when they route tickets, triage incidents, or verify agent work. For example, a support team can route by intent, a security team can prioritise alerts, and an operations team can check whether an agent action matches policy. In addition, data teams can use the model for labelling, because it can score large batches at low cost with consistent output. Therefore, the model suits teams that already have clear categories and a labelled sample to validate against.
Meanwhile, product teams that build agents should use the model as a guardrail, not as a replacement for a main model. For example, a chat agent can draft a reply, and the scorer can grade the reply for groundedness before the system sends it. However, teams should not let the scorer approve high risk actions on its own until they have set human review thresholds.
Microsoft reports similar patterns in its own work, including Xbox feedback labelling, Copilot routing, and Discovery scoring. However, those examples come from Microsoft, so external teams should treat them as patterns to test, not as promises.
Frequently Asked Questions
What is Microsoft Decision 1 pricing?
Microsoft Decision 1 pricing lists at 0.042 dollars per million input tokens, with output tokens free. For example, a system that uses 1,000 input tokens per decision can run one million decisions for about 42 dollars in scorer input cost. However, teams should confirm the rate in Microsoft Foundry at purchase time.
Where is Microsoft Decision 1 Foundry access?
Microsoft Decision 1 Foundry access runs through Microsoft Foundry, and Microsoft also offers access through OpenRouter. For example, an Azure centred team can add the model inside Foundry, while a multi provider team can call it through OpenRouter. In addition, teams should check regional availability before they route sensitive data through either path.
How does Microsoft Decision 1 latency compare with a chat model?
Microsoft Decision 1 latency looks far lower than a general chat model in Microsoft tests, with a median of about 85 milliseconds and a claim of about 35 times faster performance than GPT-6 Sol. For example, a routing step that takes seconds with a chat model can complete in under a tenth of a second with a scorer. However, real latency depends on network and region, so teams should measure it on their own traffic.
What are Microsoft Decision 1 use cases?
Microsoft Decision 1 use cases include routing, classification, prioritisation, verification, workflow control, agent guardrails, data labelling, AI judging, search relevance, and safety screening. For example, a team can route tickets, grade agent actions, or check an answer against evidence with the same scoring pattern. Therefore, the best first use case is usually one clear workflow with labelled examples and a human fallback.
What is the Microsoft Decision 1 Qwen base?
The Microsoft Decision 1 Qwen base is Alibaba Qwen3.5-9B. Microsoft post trained that base model for single pass decision scoring. For example, the base gives the model language understanding, while the post training shapes it to return calibrated probabilities for fixed options. In addition, Microsoft plans to rebase later versions on other models, but buyers should judge the current version on its own results.
Can a decision model replace a large language model?
A decision model cannot replace a large language model for open work, because it never writes prose or drafts content. For example, it can choose between options, but it cannot write the email or brainstorm a campaign. Therefore, teams should pair the two, with the large model doing open work and the scorer checking, routing, and gating the result.
Final Verdict
Microsoft Decision 1 earns attention because it solves a narrow job extremely well: it turns a fixed choice into a number that software can act on, at low cost and high speed. For example, routing, grading, and verification all become cheaper and easier to audit when the output is a probability instead of a paragraph. However, the model only works when teams define clear options and set human review thresholds. Therefore, start with one workflow, measure accuracy on your own data, and expand automation only for high confidence cases. Finally, keep the main model for open work and let the scorer guard the gates.
Independent launch coverage also tracks the Foundry release and reported benchmarks for teams that want a second source before they test.
