A model that only answers multiple choice
Microsoft made Microsoft-Decision-1.1 generally available in Microsoft Foundry on 17 November 2026 and announced it on the Foundry blog the next day. OpenRouter lists it with the same release date, and Azure is the only host behind that listing. It replaces Microsoft-Decision-1, which went into preview on 9 October.
The model has a deliberately narrow job. You send it some content and a fixed list of possible answers, and it sends back a probability for each answer. It doesn't write text or explain itself. That suits the small judgement calls inside an AI product, such as which queue a support ticket goes to, which tool an agent should call next, whether a reply needs a human to check it, or how good another model's output was. Microsoft's documentation says a generative model is still the right choice for writing, summarising or explaining, and suggests using the two together, with the decision model acting as a gate in front of the expensive one.
Microsoft says 1.1 starts from Google's Gemma 4 26B base model, adds training aimed at decision tasks, and runs on a custom inference engine.
How a request is built
Each call has a state field and one or more named questions. The state can be plain text or JSON, and it can point to images attached in an assets block. There are three question types:
- Noul answers a yes or no condition with a single probability between 0 and 1. Typical uses are "does this reply promise a refund?" or "does this document contain a prompt injection?"
- Choice picks one label from a set you define and gives a probability for every label, which makes it the tool for routing and classification.
- Score puts an item on an ordered scale you supply, from lowest to highest. The result is a probability-weighted average of the level positions, so a four-step scale returns a number from 0 to 3 that can fall between steps.
One request can hold several questions about the same state. Each question is answered on its own and can't see the others' results. For multi-select problems, the launch post recommends one Noul question per option, and it warns that those probabilities won't add up to one and shouldn't be normalised into a single confidence figure.
Requests go to a /providers/microsoft/v1/systemone path on your Foundry resource. The model field takes your deployment name. The docs recommend Microsoft Entra ID for authentication and also accept an API key, and the quickstart needs Python 3.10 or later. You can deploy it as Global Standard or, in some regions, as Data Zone Standard, which keeps inference inside one data zone.
What changed from version 1
- Image input. PNG and JPEG files go in as base64 data and are referenced in the state, the instructions or the answer options as
image:<name>. Each side of an image has to be between 280 and 1,568 pixels, and a question can render up to eight image occurrences. - Speed. Microsoft says 1.1 is more than twice as fast as its predecessor.
- Accuracy and calibration. Microsoft reports an average of 78.8% accuracy across 102 benchmarks with over 385,000 held-out questions, and says the model leads in seven of nine capability categories. These are the company's own figures.
- Evaluation judge. The model can now act as the grader in Foundry Evaluations for quality, performance and safety checks.
What it costs
Microsoft only charges for input tokens, and output is free. List prices per million input tokens are $0.042 for Global Standard, $0.046 for the US data zone, and $0.050 for the EU and APAC data zones. Classmethod's review points out that the data zone prices went up with this release while the global price stayed the same.
Artificial Analysis puts the cost of a million tasks from its test suite at $140, close to the $137 median across the 26 decision models it tracks. OpenRouter shows a median latency of about 0.15 seconds through Azure.
An early outside test
Classmethod, a Japanese engineering firm, compared versions 1.0 and 1.1 on a Global Standard deployment called from Japan. The author ran 106 scored pairs of values, each asked as an exact-match question and a semantic-match question, five times per model, alongside two competing decision models.
- Version 1.1 got 97 of 106 right, up from 95, and wrongly accepted mismatched pairs twice, down from six times.
- Version 1.0 often pushed probabilities to the extremes, with only 2 answers falling between 0.2 and 0.8. Version 1.1 put 16 answers in that range, so its confidence carries more information.
- Repeating the same requests gave identical probabilities every time on 1.1. On 1.0, results drifted by as much as 0.10.
- Measured from the client side, the median response time dropped from 1,093 ms to 1,020 ms. That's about 7% faster, well short of the 2x claim, though the author notes they couldn't separate out server time.
- On five document images with 20 deliberately planted errors, 1.1 found 17. It also flagged 18 of 61 correct values as wrong, more false alarms than one rival model and fewer than the other. It used about 759 input tokens per image on average.
- Video, audio and PDF uploads were all rejected with HTTP 400.
So the claims about calibration, consistency and image support held up in this test, while the speed claim didn't show up from the client's side.
Independent benchmarks
On the Artificial Analysis Decision Index, which subtracts a penalty for confident wrong answers from accuracy and also checks whether answers stay the same when a question is reworded, Microsoft-Decision-1.1 scores 68. That puts it 12th of 26, just above the median of 66. It's quicker than most, ranking fifth at 123 ms per task against a 199 ms median. Artificial Analysis lists the model with a 33,000-token context window and as text-only, which doesn't match Microsoft's image support, probably because the benchmark covers text tasks only.
Limits to plan around
Microsoft's documentation lists several risks. The model can carry biases from its base model and training data. Its scores can shift when you reword questions or reorder options, and a badly framed question still gets a confident-looking number. Calibration is best on familiar task types, the model gives no explanations, and its knowledge may be out of date. Used as a safety filter, it can miss subtle harmful content or flag harmless material. Microsoft says it shouldn't be the only decision maker for consequential choices about people, such as credit, employment, housing, healthcare or legal matters.
Beyond that, the 33,000-token window caps how much state you can send, only PNG and JPEG are accepted as media, and the example thresholds in the launch post are for demonstration only.
What builders can do with it
- Put it in front of a large model. Use it to route, gate or reject requests cheaply before anything reaches a costly generative model.
- Set thresholds on your own data. Collect labelled examples, then choose cut-offs based on what a false positive or a false negative actually costs you. Send the uncertain middle band to a human.
- Add a way out. Include an option like "cannot tell" so the model isn't forced to choose.
- Test how stable it is. Shuffle the option order and reword questions to see whether results hold, since both the docs and Artificial Analysis flag wording sensitivity.
- Measure from where you are. Classmethod's numbers show that client-side latency can look very different from vendor claims, so benchmark from your own region before you commit.
- Batch your questions. Asking about routing, severity and repeat contact in one request cuts the number of calls and keeps every answer based on the same context.
At these prices, a team in Auckland could run thousands of routing decisions for a few cents. The real work is in labelling enough data to trust the thresholds.
Sources
- Introducing Microsoft-Decision-1.1 (Microsoft Foundry Blog)
- Deploy and use Microsoft-Decision-1.1 in Microsoft Foundry (Microsoft Learn)
- Microsoft-Decision-1.1 (Artificial Analysis)
- Microsoft-Decision-1.1 API Pricing and Benchmarks (OpenRouter)
- Testing Azure Microsoft-Decision-1.1 (DevelopersIO, Classmethod)
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.