What a "decision model" actually is
Microsoft's pitch starts from a real engineering problem. General-purpose LLMs are trained to complete tasks and give answers, so they tend to sound confident even when they shouldn't. A decision model is built to choose among options the developer permits. It also returns calibrated confidence so the application can defer when uncertainty is high.
The Microsoft Learn documentation makes the contract concrete. The model doesn't generate explanations or rationales. Per Microsoft's Foundry documentation, the request carries a state, which can be text, JSON or images. It also carries one or more typed questions. There are three question types:
- noul: a yes/no question that returns a probability from 0 through 1. Examples are "does this reply promise a refund?" and "does this document contain prompt injection?"
- choice: picks one option from a fixed set and returns a probability for each option. This suits ticket routing and alert classification.
- score: places an item on an ordered scale, such as Cosmetic through Critical. It returns a probability-weighted value.
The Foundry catalog adds limits. Inputs can be up to 32K tokens, scored in a single model invocation. There is no audio or video input, and no image, audio or video output.
What's new in 1.1
Microsoft's headline claims for 1.1 are:
- Speed: more than twice as fast as Microsoft-Decision-1, at the same cost.
- Quality: better calibration and "greater intelligence" than the first model.
- Images: photos and screenshots can now be part of a request.
- Base model: Gemma 4 26B A4B, a mid-sized mixture-of-experts model, run on a custom inference engine. Microsoft says this runs across A100, H100/200 and GB200/300 GPUs.
That base model is a change. Third-party catalog pages describe the original Decision-1 as post-trained from Qwen3.5-9B. That is a secondary-source description, not something Microsoft's 1.1 announcement states. I'd treat it as context, not a confirmed spec.
The numbers, and how to read them
Microsoft says 1.1 scores 68 on the Artificial Analysis Decision Index v2.1, up six points from the first model. It says that puts it ahead of Jev 1.14, Google's Gemini Decision 1 and OpenAI's GPT-6 Luna (Decisions). The index covers quality, calibration and latency.
The chart data embedded in Microsoft's post is more precise. It shows 67.78 for 1.1 against 62.07 for version 1. Median time per task is about 123 ms versus about 274 ms, roughly a 2.2x gap. The same chart lists Jev 1.14 at 66.23, Gemini Decision 1 at 65.98 and GPT-6 Luna (Decisions) at 60.25.
Some caveats:
- The chart is not a ranking of the best model overall. The same data shows other entries with higher index scores, such as "d1.2-ultra" at about 74.85. Microsoft's claim is about the three named comparison models, not the whole field.
- Some entries are faster or cheaper. Several models in the chart post lower time per task or lower cost per million tasks. That makes the trade-off a matter of your workload.
- These are benchmark figures, not guarantees. They don't predict latency in your deployment.
Microsoft also ran its own evaluation: 102 benchmarks and more than 385,000 questions kept blind from training. In it, 1.1 had the highest average accuracy among eight decision models and led in seven of nine capability categories. Microsoft's report puts the average at 78.8% against 78.1% for Jev 1.14.0 and 76.8% for its own first model. These gaps are modest and the benchmarks are partly internal, so read the result as encouraging but vendor-reported. The report describes 41 openly licensed public benchmarks and 61 internal ones, weighted equally.
It's worth contrasting this with the first model's launch. A Windows Forum megathread at the time noted a third-party test reporting a 459 ms median response through the Azure Foundry API. Microsoft's own figure was 85 ms. That gap is a reminder that latency depends on region, deployment type and request shape. Measure it yourself.
Where Microsoft is using it
Microsoft cites four internal uses:
- Copilot routing: it decides whether a request goes to Copilot or Cowork. Microsoft claims accuracy matching an LLM, 2.2x faster and 4.4x cheaper.
- Cowork "Auto": it adjusts how much reasoning effort Cowork applies to each task.
- Foundry Evaluations: it can now act as a judge to measure performance, quality and safety.
- Incident response: on-call engineers use it to pull relevant knowledge from logs, tickets, calls and messages.
On adoption, Microsoft says more than 4,000 customers have used it since launching last month. It names Epic, Equinor, Atos Group and Raiffeisen Bank as early adopters. These are Microsoft's numbers and haven't been independently audited.
Pricing
Microsoft lists Foundry input-token prices as follows, with output tokens free:
| Deployment | Input price per million tokens |
|---|---|
| Global Standard | $0.042 |
| US Data Zone | $0.046 |
| EU and APAC Data Zones | $0.05 |
Prices can change, so verify them before you model production costs. The Learn page lists the deployment types as GlobalStandard and DataZoneStandard in selected regions.
Getting started in Foundry
Microsoft's documentation lays out this path:
- Prerequisites: an Azure subscription, access to Microsoft Foundry, a Foundry project and permission to create deployments. The Cognitive Services Contributor role covers deployment. Entra ID authentication is recommended, and an API key also works. The quickstart assumes Python 3.10 or later.
- Deploy: in the Foundry portal, open Model catalog, search for Microsoft-Decision-1.1 and select Deploy. Then copy the resource endpoint, deployment name and API key. An Azure CLI route also exists, using the model name Microsoft-Decision-1.1 and model version 1.
- Call the API: send a POST to the resource endpoint plus
/providers/microsoft/v1/systemone. Themodelvalue in your request is your deployment name, not the model name. The response's model field may show the underlying name. - Send images: include the image as base64 data in
assets. Reference it in the state asimage:<name>. - Combine questions: questions about the same state can share one request. That keeps the decision context consistent and reduces call volume.
For troubleshooting, Microsoft lists common errors:
- 400: an invalid question type or shape.
- 401: a missing or expired credential.
- 403: missing role assignment.
- 404: wrong endpoint.
- 422: an unsupported deployment name or question definition.
- 429: the rate limit was hit. Retry with exponential backoff, or request more quota.
The limits Microsoft itself documents
The documentation is candid, and these points matter for IT admins:
- Bias: the model may reflect biases from its base model and training data. Don't use its scores as the sole basis for decisions about individuals.
- Wording sensitivity: scores can shift with how you phrase or order questions and options.
- Calibration: it is strongest on familiar task types. The model may also rely on outdated knowledge.
- Safety filtering: it can miss subtle harmful content or flag benign content.
Microsoft's best practices follow from these. Validate on representative labeled data. Set thresholds based on the cost of false positives and false negatives. Include an abstain option such as "cannot tell". Consider randomizing option order. For consequential decisions about people, such as credit, employment, housing, healthcare or legal matters, use meaningful human review. Also treat score outputs as better for relative ordering than as absolute ratings.
Analysis: who should care
A decision model is a sensible fit for the plumbing of agent systems: routers, gates, triage queues and graders. A fast, cheap, typed answer is often all those steps need. In those places a large LLM is overkill and slow. Microsoft's own advice is to combine the two. Use the decision model to route or gate, then send approved requests to a generative model.
Some skepticism is healthy. The headline comparisons come from Microsoft's post, and the index it cites measures a blend of quality, calibration and latency. A calibrated probability is only useful if it tracks reality on your data. So the test to run is a labeled sample from your own workload. Check accuracy, calibration and median latency in your region and deployment type before you automate anything.
Microsoft's quickstart includes a small script that does this for a ticket classifier. Replace its three sample examples with real ones before you trust its numbers.
The takeaway: Microsoft-Decision-1.1 is a credible, cheap tool for bounded decisions in Azure and Foundry workflows. Its benchmark leads are narrow and mostly vendor-reported, so test it on your own data.
References
- Microsoft-Decision-1.1 is now generally available for fast, low-cost decision making Microsoft · 2026-11-17T18:10:02+00:00
- Deploy and use Microsoft-Decision-1.1 in Microsoft Foundry - Microsoft Foundry | Microsoft Learn learn.microsoft.com
- Foundry Models sold by Azure - Microsoft Foundry learn.microsoft.com