Microsoft-Decision-1 is a Microsoft model for scoring predefined choices in software workflows. Announced on 9 October 2026, it targets classification, model routing, response evaluation, and agent control.[1]

Model overview

Property	Details
Publisher	Microsoft.[2]
Base model	Qwen3.5-9B, post-trained by Microsoft.[2]
Input	Text.[2]
Output	Structured probability scores, rather than generated prose.[2]
Context window	32,768 tokens.[2]
Foundry lifecycle	Generally available, version 1.[2]
OpenRouter identifier	`microsoft/microsoft-decision-1`.[3]
Listed token price	US$0.042 per million input tokens; output tokens are free.[3]

Decision scoring

An application supplies context, a closed question, and answer options. The model scores the options in one invocation. It supports binary questions, multiple-choice questions, and ratings, including grading against a supplied rubric. Training combines public datasets reviewed through Microsoft's Open Data process with synthetic examples.[2]

The model card permits abstention options for insufficient evidence and assigns the integrating application responsibility for confidence thresholds, human review, and safeguards around subsequent actions.[2]

Access and API formats

Microsoft offers the model through Foundry. OpenRouter lists Azure as its provider under microsoft/microsoft-decision-1, with the same published input price and free output.[3]

Vercel added the model to AI Gateway on 9 October 2026. Gateway users can access it without a separate Azure account or deployment, using the AI SDK decision API, the OpenAI-compatible Decisions API, or the TypeSafe-compatible API. Several questions can use the same input within one request.[4]

In the AI SDK, experimental_decide takes model, state, and questions. Boolean answers contain a probability. Choice answers contain the selected option and probabilities; score questions use an ordered rating scale. Gateway's HTTP equivalent is POST /v1/evaluate. These decision calls are distinct from Chat Completions and Responses requests.[5]

The following illustrative request defines a binary check; it does not prescribe the probability the model will return.[5]

import { experimental_decide as decide } from 'ai';

const result = await decide({
  model: 'microsoft/microsoft-decision-1',
  state: 'The test process finished with exit code 0.',
  questions: {
    succeeded: {
      type: 'boolean',
      instructions: 'Did the test process succeed?',
      criteria: {
        true: 'The exit code is 0.',
        false: 'The exit code is nonzero.',
      },
    },
  },
});

Confidence calibration

Calibration concerns whether confidence matches observed correctness, rather than whether the model chooses the right answer overall. For a calibrated classifier, predictions assigned 80% confidence should be correct about 80% of the time. Guo and colleagues demonstrated that classification accuracy and probability calibration can diverge in neural networks.[6]

Microsoft-Decision-1's probability output does not include an explanation or rationale. It also does not process images, audio, or video.[2]

Microsoft evaluations

Microsoft reported the highest accuracy among its compared models across 36 benchmarks comprising nearly 150,000 questions, withheld from training. Its median-latency comparison reported a 35-fold speed advantage over GPT-6 Sol. In eight types of input perturbation, decisions changed on average in 1.3% of cases. These are Microsoft's measurements for the tested tasks.[1]

Internal trials included sorting over 10,000 items of Xbox feedback, assessing Copilot responses, incident-related retrieval, and scoring experiments for Microsoft Discovery's replanning process. Microsoft described these as internal tests.[1]

Intended boundaries

Microsoft excludes open-ended conversation, translation, summarization, and tasks requiring information absent from the input. Its model card also excludes sole automated use in consequential decisions about people, including employment, credit, housing, healthcare, and legal rights.[2]

References

Achint Srivastava, Microsoft Command Line. Introducing Microsoft-Decision-1, our model for fast decision-making, 9 October 2026.
Microsoft Foundry. Microsoft-Decision-1 model catalog, accessed 10 October 2026.
OpenRouter. Microsoft-Decision-1 API pricing and providers, accessed 10 October 2026.
Vercel. Microsoft Decision-1 now available on AI Gateway, 9 October 2026.
Vercel. AI Gateway decision documentation, accessed 10 October 2026.
Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On Calibration of Modern Neural Networks, ICML 2017, PMLR 70:1321-1330.

Editing: Microsoft-Decision-1