Guide Microsoft AI/SI · Chapter 2

    Which AI model does which work at Microsoft.

    GPT, Claude, Grok and open models sit in the Microsoft Foundry catalogue, under Copilot a router chooses. Which model takes on a task is decided by your test set.

    As of: October 2026 · Editor: Marcel Haas

    • Microsoft Foundry
    • Azure OpenAI
    • Model Router
    • Claude
    • Grok
    • Ollama

    Model data sheets

    One data sheet per model family.

    For every family: who offers it, where it is available, where it computes, what we test it on first, and the Stand with the vendor's page where one exists.

    • OpenAI, in Foundry as Azure OpenAI

      GPT-6 family

      Where it is available
      Microsoft Foundry, under Copilot through the router, directly through the OpenAI API
      Where it computes
      Per version Standard, Data Zone Standard or Global Standard; the Swiss region carries GPT-4.1, GPT-4o and the embedding models as of September 2026
      What we test it on first
      Office work, coding and computer use with Astra and Sol; classification, extraction and short texts at volume with Luna
    • Anthropic, in Foundry as a partner model

      Claude 5 family

      As of 21 September 2026

      Where it is available
      Microsoft Foundry through Azure Marketplace, under Copilot and in Copilot Studio after approval, directly at Anthropic
      Where it computes
      Global Standard, hosted on Azure or by Anthropic; a data zone for some models in the US only, no regional deployment
      What we test it on first
      Long contracts and reports, analyses with references, development work with Claude Code; Haiku for small, fast steps
    • SpaceXAI, in Foundry as Foundry Models sold by Azure

      Grok

      As of 5 October 2026

      Where it is available
      Microsoft Foundry with sign-in through Entra ID or an API key, directly from the vendor
      Where it computes
      Global Standard with global processing; a data zone in the US only, no regional deployment
      What we test it on first
      Research with current sources and market monitoring, when freshness matters more than citation fidelity; results are checked against your sources
    • Mistral AI, Meta, Microsoft, DeepSeek; under each maker's licence

      Open models: Mistral, Llama, Phi, DeepSeek

      Where it is available
      Microsoft Foundry, locally or on GPU machines in Azure through Ollama
      Where it computes
      In Foundry per model and deployment type; with Ollama on your hardware, with no external requests
      What we test it on first
      Classification, summarisation and translation at volume when cost and control count; Phi for small tasks at the edge

    The families we test with first; your test set decides. A card without a Stand names the model as the catalogue lists it until the vendor's page is in the index.

    01

    A model reads the context and returns an answer, a tool call or a structure.

    A clerk attaches the minutes of a meeting and asks for the resolutions with owner and deadline. For that, the model sees everything in the context window: the instruction, the minutes, the description of the tools and the schema of the desired output, counted in tokens, and from it writes an answer, a tool call or an output in exactly the prescribed shape. A reasoning model thinks in steps first and needs more tokens and time for it; a small model answers quickly and cheaply as long as the task stays narrow. Context, reasoning, tool calls and structured output decide whether a model carries a task, and all four can be measured on your cases.

    02

    The Foundry catalogue carries the families we test with first.

    In the Microsoft Foundry model catalogue you create every model as a deployment in your subscription: the GPT-6 family as Azure OpenAI, with Astra and Sol for office work, coding and computer use and Luna for classification, extraction and short texts at volume; the Claude 5 family from Anthropic as a partner model through Azure Marketplace, for long contracts, reports and analyses with references; Grok from SpaceXAI for research with current sources; and the open models Mistral, Llama, Phi and DeepSeek for classification, summarisation and translation when cost and control count. The same open models run with Ollama on your own hardware. Azure Policy limits which models may be deployed in a subscription at all, and because every family sits in the same catalogue, a later switch stays open.

    03

    Under Copilot a router chooses, in Foundry you choose or the Model Router does.

    Whoever works with Copilot in Word or Teams picks no model: in Auto mode Copilot decides per request whether a fast model is enough or one for deeper reasoning is needed, and the person can switch themselves. Models from Anthropic sit next to it as a subprocessor, in the EU, the EFTA and the UK only after approval. In Copilot Studio you set the primary model of an agent, since June 2026 also Claude Sonnet 5 or GPT-5.5 Chat. In Foundry either your application decides per call, or the Model Router as a deployment of its own sends simple requests to small models and hard ones to large models, within the set you approve. The router saves tokens where many simple requests arrive, and it needs the same test set as a fixed model, because the answer to each request may come from a different model.

    04

    The deployment type decides where a model computes.

    A model in the catalogue is not yet a place. Only the deployment in your subscription sets it: Standard computes in the chosen Azure geography, Data Zone Standard within the data zone, Global Standard in any Azure region with free capacity; added to that are reserved capacity for even load and batch for jobs without hurry. For Switzerland North, as of September 2026, GPT-4.1, GPT-4o and the embedding models run as a regional deployment; newer GPT models process in the EU data zone or globally, Claude and Grok globally or in the US data zone. That is why the data sheet of every deployment carries model, version, deployment type and region, and the data boundaries chapter says which data class may go to which row.

    05

    The test set decides, and retirements belong in the plan.

    Two candidates work through the same set of real cases from your business team: correct, incomplete and forbidden requests, documents with hidden instructions, outages. We compare answer quality, fidelity to the passage, speed, the review effort for people and the cost per case, and the proposal records why one model won; Foundry measures groundedness and tool call accuracy before every new version and continuously in operation, and after every change to model, instructions or sources the test set runs again. A model does not stay forever: Microsoft retires model versions in Azure OpenAI with a date, and Prompt flow in Foundry is being retired. Every data sheet therefore carries the date on which the version is replaced, and the test set is the tool with which the successor takes over without surprises.

    What you decide.

    Which model for the first case?

    Options

    1. 1The model the business team knows from daily work
    2. 2The strongest model in the catalogue
    3. 3Two candidates on the test set, decided by result and place

    Our recommendation: Two candidates on the test set. For classification, extraction and summarisation a small model competes against a medium one, and often the small one is enough, deployed regionally; for long contracts, analyses with references and coding a large GPT model competes against Claude. The winner is the one that scores better on the test set and may compute in the place the data class demands.

    What we advise against: Choosing the model by ranking or headline. A vendor's benchmarks measure someone else's tasks, your test set measures yours.

    Model Router or a fixed deployment?

    Options

    1. 1One fixed deployment per application
    2. 2The Model Router with an approved set of models
    3. 3Both: fixed for answers with consequences, the router for the volume

    Our recommendation: Fixed where an answer has consequences or the place must be certain: then it is traceable which model answered, and the region is set. The router where many simple requests arrive and token consumption counts, with a set that contains only models in the permitted place, and with the same test set as a fixed model.

    What we advise against: Switching on the router to postpone the model choice. It chooses per request, and without a test set nobody knows whether the small models in the set are up to the task.

    One model or several per application?

    Options

    1. 1One model for every step
    2. 2A main model and a small model for preparation and follow-up
    3. 3Plus a second model as a checker

    Our recommendation: A main model for the task and a small one for classification, extraction and pre-sorting, because the bulk of the calls happens there. A second model as a checker pays off where a wrong answer is expensive: it checks references and completeness before a person decides. Several model families are at the same time the answer to the insider risk of a single model.

    What we advise against: Several models because the catalogue has them. Every additional model needs a data sheet, a place and a slot in the test set.

    When do we switch the model?

    Options

    1. 1When the vendor retires a version
    2. 2When a new model scores better on the test set
    3. 3At a fixed date every quarter

    Our recommendation: When the test set shows it or the retirement date demands it, and never unnoticed. Every data sheet carries the date of replacement; a quarter before, the successor competes on the test set, and the switch goes through evaluation and approval as a new version, like every other change to instructions or sources.

    What we advise against: Moving to the newest version because it is there. A new version answers differently, and what has proven itself in instructions and sources must be measured again.

    Status and sources.

    As of: October 2026

    • As of 21 September 2026

      Claude runs in Foundry, globally or in the US data zone.

      Claude models are partner models through Azure Marketplace, hosted on Azure or on Anthropic's infrastructure. All versions run as a Global Standard deployment; a data zone exists for some models in the US only. Authentication uses Entra ID or an API key; the Mythos models accept Entra ID only.

      Source: Microsoft Learn: Deploy and use Claude models in Microsoft Foundry
    • As of 5 October 2026

      Grok 4.7 is generally available in Microsoft Foundry.

      Microsoft sells the Grok models as Foundry Models sold by Azure. Grok 4.7 accepts text and images, calls tools and supports the Responses API; authentication uses Entra ID or an API key. Live search options do not apply to deployments on Azure.

      Source: Microsoft Learn: Deploy and use Grok models in Microsoft Foundry
    • Since June 2026

      New authoring experience, new primary models.

      The new authoring experience on the GitHub Copilot harness is available as a production-ready preview: Microsoft IQ for emails, calendar, files and Teams messages, reusable skills and memory per user. Claude Sonnet 5 and GPT-5.5 Chat are generally available as the primary model.

      Source: Microsoft Learn: What's new in Copilot Studio

    Frequently asked questions.

    Do the same GPT models run in Foundry as at OpenAI?

    Yes, as Azure OpenAI: the same model families, deployed in your Azure subscription, signed in with Entra ID, billed on the Azure invoice and logged in Purview. Which version is available in which region and with which deployment type is listed per model in the catalogue, and the data sheet of the deployment records it.

    What exactly does the Model Router do?

    It is a deployment in Foundry itself. It inspects every request and forwards it to one of the models you have approved in its set: simple ones to small models, hard ones to large models. In the trace you see which model answered, and Azure Policy limits which models may enter the set. For tasks where the place must be certain, the set contains only models with the matching deployment type.

    Can employees pick a model in Copilot?

    In Auto mode Copilot chooses per request, and the person can switch to a fast model or to deeper reasoning. Whether models from Anthropic are among them is decided by your IT: in the EU, the EFTA and the UK they are enabled for no users by default and excluded from the EU Data Boundary. For an agent in Copilot Studio, the person who builds it sets the primary model.

    How do we know that a model carries the task?

    By the test set: real cases from your business team with the expected answer, among them incomplete and forbidden requests. We measure answer quality, fidelity to the passage, tool calls, speed, review effort and cost per case, and only a model that scores no worse there than its predecessor goes into the new version. Foundry measures the same quantities continuously in operation.

    What happens when Microsoft retires a model?

    The version gets a retirement date, which is recorded in the data sheet. Before that, the successor competes on the test set, instructions and sources are adjusted where the answers differ, and the switch goes through evaluation and approval as a new version. Deployments nobody updates expire on the retirement date; that is why one person per application maintains the data sheet.

    When is an open model on your own hardware worth it?

    When text may not leave the environment, when many small calls arise that a small model carries, or when a team wants to test with sample data before creating a subscription. With Ollama the same model runs locally or on a GPU machine in your Azure subscription, with OpenAI-compatible endpoints. Quality, memory needs and operation we check on the test set like for any other model, the model licence per use.

    Next chapter · Chapter 3

    What an AI agent at Microsoft is made of and where it runs.

    What is an agent made of at Microsoft, and how does it run safely?