# Microsoft Decision 1.1: Pricing, API Access and When to Use It

Decision 1.1 scores fixed answers instead of generating text. Here is how to evaluate its $0.04-per-million-input-token offering for routing and classification.

By Paras Madan | 2026-11-18


Microsoft Decision 1.1 is a useful model to consider when the answer should come from a defined set of options, not a paragraph. Think of routing a support request or checking whether a response satisfies a rubric. In those cases, generating an explanation and then parsing it can add work your application does not need.

The practical question is whether a compact scoring step can replace that part of your workflow without increasing mistakes. I would evaluate it on a bounded classification problem before putting it in front of an agent's actions.

## What Decision 1.1 actually returns

Microsoft's [Foundry catalog](https://ai.azure.com/catalog/models/Microsoft-Decision-1.1?publisher=microsoft) marks Decision 1.1 generally available. It returns probabilities for supplied answer options rather than generated text, with support for formats including yes/no, multiple choice and ratings.

The catalog lists approximately 32K context and up to eight input images. It is built on Gemma 4 26B-A4B and post-trained by Microsoft. It does not generate explanations, accept audio or video, or replace open-ended conversation and summarization.

That distinction matters when choosing the interface. If the application needs a written reply, this model is not the component that writes it. It can instead help decide where the request goes or whether another model's reply meets specified criteria.

## What the published price means

The [Vercel AI Gateway model page](https://vercel.com/ai-gateway/models/microsoft/decision-1.1) lists $0.04 per million input tokens. Its example uses the model identifier `microsoft/decision-1.1` with AI SDK's `experimental_decide`, passing a state and structured questions.

For a hypothetical one billion billed input tokens, that rate produces $40 in model input charges. That is arithmetic, not a forecast for one billion decisions. The number of requests you can serve depends on how much input each request contains.

Keep that distinction in a budget. Measure billed input for representative cases, including the actual instructions and material you submit. Do not infer a fixed price per classification from the token rate. Also keep gateway, application and downstream model costs separate where they apply.

## Define the answer space before the threshold

Microsoft explicitly supports an abstention option such as “cannot tell.” The integrating application remains responsible for its thresholds, escalation and safeguards. The model is not intended as the sole automated decision-maker for consequential decisions about people.

For an ordinary support-routing test, I would define billing, technical support and account access alongside an unclear or unsupported category. Give each label a precise meaning and include examples that sit between categories.

Without that work, a high score can make an ambiguous classification look more definite than it deserves. A probability is useful only in relation to the question and options you supplied.

## Test calibration on your own examples

Build a small labeled dataset from requests whose correct destination is already known. Keep a separate held-out set so you are not repeatedly adjusting the prompt against the same examples and calling the improvement general.

Measure routing accuracy, but also inspect confidence. Among predictions near a given probability, how often is the selected answer actually correct? That is a practical check of calibration on your workload, rather than an assumption based on the model description.

Then compare different escalation thresholds. Track the mistakes that escape review and the extra work sent to a person. The right balance depends on the cost of each kind of error, not on finding a universal confidence number.

## Run in shadow mode before changing the workflow

Initially, record what Decision 1.1 would have selected while keeping the current routing system in charge. Review disagreements and determine whether they reveal a model error, an unclear label or an outdated historical answer.

Measure latency end to end, not just model response time. Include network calls and any fallback. A fast scorer is less useful if an unreliable integration turns routine requests into retries.

Keep the first production use reversible. A bounded routing change with a clear fallback is easier to assess than a model that immediately gains authority to execute actions.

## Use scoring where scoring is enough

Decision 1.1's appeal is its narrower job. It can return a structured choice without pretending to be the assistant that handles the entire request.

The strongest evaluation is therefore specific: can it make this choice accurately, with useful uncertainty and acceptable operating cost? Answer that before expanding it into more workflows.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain technical products through practical developer stories.

[Book a 30-minute call](https://cal.com/paras-madan/30min).

