Microsoft's Decision-1 prices a decision at $0.042 per million input tokens. Calibration is what you buy

How sure is your router?

Most agent systems cannot answer. A generative model asked to pick a category returns a word, and the word comes with no reliable sense of how close the call was. Microsoft-Decision-1, which Microsoft put into public preview in Foundry on October 9, is built to answer that question first.

Given a fixed set of options, the model returns a calibrated probability for each one. It handles yes or no, multiple choice and rating questions, plus rubric-based grading of AI responses and agent actions, through a structured API. Microsoft lists routing, classification, prioritization, verification and workflow control as the uses. It is built on Qwen3.5-9B, post-trained for single-pass decision scoring, and priced at $0.042 per million input tokens in both the US and EU Datazones, with output tokens free. It is also available through OpenRouter.

Three vendors, one layer

This is the third product in the same category we have covered since October 5. Cloudflare argued for decision models on the agent’s hot path with Clef, and OpenAI priced the layer separately with its Decisions API. The case for separating small, frequent decisions from expensive generative reasoning no longer needs making. What differs now is how each product behaves.

Decision-1’s distinguishing claims, all reported by Microsoft:

• Probabilities, not labels. A score per option lets a workflow set thresholds: act above one value, escalate below another.
• Speed and robustness. Microsoft reports the highest accuracy in a 36-benchmark comparison of nearly 150,000 questions, a P50 latency about 35 times faster than GPT-6 Sol, and decision changes on 1.3% of input perturbations, with none for reordered or paraphrased options.
• An open-weight base, which matters for anyone who later wants to run the same behavior elsewhere, and a published price where only inputs are billed.

What you are actually buying

The price is low enough that cost will rarely decide this. What you are buying is calibration: the promise that when the model says 0.9, it is right about nine times in ten. Microsoft describes calibration as holding “on representative cases.” Yours may not be representative.

Calibration is cheap to test and expensive to assume. Take a few hundred labeled examples from your own workflow, bucket the model’s predictions by confidence, and check accuracy per bucket. If the 0.9 bucket is right 70% of the time, your thresholds are wrong, however good the benchmark looks. Run the same check on each option, because a model can be well calibrated on common classes and badly calibrated on rare ones, which are usually the ones that matter in escalation.

Test the robustness claim the same way. Reorder your options, rephrase their descriptions, add an irrelevant option. A decision that changes is a decision you cannot audit.

A model that will change underneath you

Microsoft says it will soon rebase Decision-1 on other models, including its own MAI models and OpenAI’s. That is a product roadmap, and it is also a warning: the behavior you calibrate in October may not be the behavior you get after the rebase. Pin the version you validated, and treat every rebase as a model change that reruns the calibration check before it reaches production.

There is a governance gain here too. A decision model produces a compact record: the options offered, the probability of each, the threshold applied and the path taken. Logged per decision, that is far easier to audit than a paragraph of generated reasoning. In the model portfolio we expect enterprises to run, the decision model may become the most reviewed component, because it is the one deciding where everything else goes.

Decision-1 is in preview and its numbers are Microsoft’s. The architecture it represents is not in doubt. Whether its confidence deserves your trust is a question only your data can answer, and the test takes an afternoon.