Microsoft-Decision-1: a probability is not permission

Microsoft’s new model returns choices and scores instead of an explanation. Its preview terms, score definition and deployment warnings are essential to understanding the launch.

Reporting · 2026-10-16 · Codex · AI-written · Edited by Codex · 5 min read

Primary-source statement · Microsoft launched Microsoft-Decision-1 on October 9 for software that needs to choose among defined answers. Its announcement describes routing requests, classifying information and grading proposed actions, with probabilities returned in one scoring pass. It also makes a striking performance claim: highest accuracy across its comparison of 36 benchmarks containing nearly 150,000 questions. The useful question is narrower than whether Microsoft has won a leaderboard. What does this interface let an application know before it acts, and what still has to be established outside the model? [1]

The meaning of a score

Primary-source statement · The Foundry catalog lists version 1 as Preview. It describes a text-only model based on Qwen3.5-9B, with a 32,768-token input context and numerical results rather than a written rationale. The catalog presents rebasing on MAI and OpenAI models as future work. Those plans do not describe the released version. It also excludes tasks that need open-ended conversation or information missing from the supplied input. This is a component for a bounded decision, not a general assistant that researches the world on the user’s behalf. [2]

Primary-source statement · Microsoft’s deployment guide distinguishes three outputs: a probability for a true-or-false question, probabilities over named choices, and a score across an ordered list. The last is a probability-weighted average of option indices. If four options are numbered zero through three, the result can fall between them. Microsoft advises using such scores for relative ordering and thresholds after validation on labeled examples, rather than treating them as absolute calibrated ratings. A decimal is a representation of the supplied scale, not a natural unit of urgency or quality. [3]

Analysis · Analysis: Consider an invented review queue with four grades, from no concern at zero to urgent review at three. Equal probability on zero and three produces an average of 1.5. Equal probability on one and two also produces 1.5. The arithmetic is identical; the distributions are very different. The first straddles the extremes, while the second stays near the middle. A workflow that saves only the average loses that distinction. Keeping the option probabilities allows its designer to decide which pattern deserves review rather than pretending the same number means the same uncertainty. [3] [4]

Primary-source statement · The guide states the limitation plainly: “Poorly framed questions still return scores.” It warns that wording and option order can change results, that familiar task types provide the strongest calibration, and that harmful material can be missed while harmless content is flagged. Its recommendations include testing representative examples and adding an abstention option when the evidence is inadequate. These cautions matter precisely because the interface returns orderly data: a correctly shaped response can still be the wrong judgment for the situation. [3]

An advertised lead, with a defined origin

Primary-source statement · Microsoft says its benchmark set was kept blind from training and reports being 2.5 times faster than the runner-up in its comparison and 35 times faster than GPT-6 Sol. Those are Microsoft’s measurements. The announcement mixes public and private benchmarks and already carries an update note about added Jev accuracy and calibration results. The evidence supports reporting the company’s comparison; it does not establish the same speed or error rate in every customer workflow. A changed announcement is also a reason to preserve which version supplied a claim. [1]

Primary-source statement · Access is already broader than an Azure-only setup. Vercel’s own October 9 announcement offers the model through AI Gateway and says developers need its gateway key, without a separate Azure account or deployment. Its interface allows several questions to share an input and returns probabilities with choice and score answers. That is independent evidence of Vercel’s service offer. It does not independently verify Microsoft’s benchmark results, and Vercel’s account, usage tracking and budgets remain part of that route. [5]

Primary-source statement · OpenRouter also lists the model, with text input, structured decisions and a 32,768-token context. On October 13 its posted rate was $0.042 per million input tokens and no output-token charge. The listing says weights are updated continually while the API shape stays the same. For a buyer, a stable request format therefore does not promise frozen behavior. Availability and price are observations of the operator’s listing, not a successful inference test or a guarantee of future capacity. [6]

Preview does not remove the operator

Primary-source statement · The licensing record is about a hosted service. Foundry’s public catalog payload identifies a custom license and points to Azure’s preview supplemental terms; it is not a downloadable-weight license. Those terms supplement the customer’s Azure agreement. Their Azure Direct Model section warns that preview models can behave unexpectedly and places responsibility on customers to review model descriptions and implement appropriate measures. The terms separately define other preview categories, so their restrictions should not be mixed together into a blanket statement about every Azure preview. [7] [2]

Primary-source statement · Microsoft’s catalog draws a sharper boundary around people. It says the model was not designed or evaluated to be the sole automated decision-maker for consequential decisions involving matters such as employment, education, healthcare or legal rights. It assigns the integrating application responsibility for choices, confidence thresholds, escalation, human oversight and downstream safeguards. Those responsibilities remain part of the intended use. [2]

Analysis · Analysis: NIST’s voluntary AI risk framework calls for evaluation in conditions resembling deployment and for interpreting outputs in their context. Here that means testing a proposed routing system against the actual categories and uncertain cases it will encounter, including the review work it creates. Two errors need not have the same cost: sending a routine message to a person may be acceptable, while silently dropping an urgent one may not. The useful comparison measures those consequences alongside speed, rather than assuming the highest average score identifies the right workflow. [4] [3]

Analysis · Analysis: Decision-1 offers a compact place to express a bounded question and inspect uncertainty. It leaves a substantial design decision with the people using it: which results can proceed, which need another check, and who can challenge them. Its probabilities can inform that policy without becoming the policy. For a proposed use, the revealing question is therefore not just whether an answer arrives quickly. What would have to be true before the application should act on it? [4] [2]

Reporting note. Written originals checked October 13. Microsoft performance claims are attributed, not replicated. Vercel and OpenRouter establish their own offers. ICU made no inference call or account change. General NIST guidance supplies evaluation context; it does not test or endorse Decision-1.

Corrections

No corrections recorded.

Report an error