Deploying an LLM to Azure AI Foundry project.

By Ivana Tilca · September 12, 2025 · 6 min read

A complete, beginner-friendly walkthrough for deploying a large language model like GPT-5 in Azure AI Foundry — from creating the resource and picking the right deployment type to running your first prompt in the Chat Playground, plus the quota and cost gotchas nobody warns you about.

If you're just getting into building with AI on Azure, the first real milestone is deploying your own model. Not calling a shared demo endpoint — actually standing up a deployment that belongs to you, that you control, and that you can wire into your own apps. In this guide I'll walk you through the whole thing in Azure AI Foundry, from an empty portal to chatting with your own GPT-5 deployment. No prior experience assumed.

Azure AI Foundry (you'll find it at ai.azure.com) is Microsoft's unified platform for building AI applications. Think of it as the place where you discover models, deploy them, test them, and then connect them to your code — all in one resource instead of stitching together five different Azure services like we used to.

Before you start: what you'll need

A few things need to be in place first, and getting them right now saves a lot of frustration later:

An Azure subscription. A free or pay-as-you-go account both work. Some frontier models require a pay-as-you-go subscription rather than free credits.

An Azure AI Foundry resource and a project. The resource is the top-level container (billing, security, region); the project is your workspace inside it. If you don't have one yet, create it first — the portal will also prompt you to create one the first time you land on it.

The right permissions. You need at least Contributor (or the Azure AI Developer role) on the resource to deploy a model. If you're using a company subscription, this is the step that most often needs an admin's help.

Why "deploying" is its own step (and what it actually does)

This trips up a lot of newcomers, so it's worth thirty seconds. In Foundry there's a difference between a model (the raw capability sitting in the catalog) and a deployment (a live, addressable instance of that model with its own endpoint, its own quota, and its own name). When you deploy, Azure provisions capacity for you and gives you an endpoint URL and an API version. That endpoint is what your app will call. So "deploying" isn't downloading anything — it's telling Azure "reserve me an instance of this model that I can send requests to."

Step 1 — Open the portal and select your project

Go to ai.azure.com and sign in. On the landing screen you'll see your AI Foundry resource and the project underneath it. Click into the project — everything you deploy lives inside a project, not at the resource level.

Step 2 — Open the Model Catalog

In the left-hand menu, choose Model catalog. This is the storefront of everything you can deploy: OpenAI models (GPT-5, GPT-4o and friends), plus open models from Meta, Mistral, Microsoft's own Phi family, and many more. Use the search box to find what you want. For this walkthrough I'm deploying GPT-5.

Tip: the catalog shows useful metadata on each model card — context window, modality (text, vision, etc.), and hosting option. It's worth reading before you commit, especially if cost or latency matters for your use case.

Step 3 — Deploy the model

Open the GPT-5 card and click Use this model, then Create resource and deploy. Azure will ask you to confirm a deployment name and a deployment type. Don't rush past the deployment type — it's the choice that affects your cost and your rate limits the most:

Standard — pay per token, hosted in a specific region. A good default for most apps and for learning.

Global Standard — pay per token, but routed across Microsoft's global capacity, which usually means higher throughput limits. Great when you expect traffic.

Provisioned (PTU) — you reserve dedicated throughput for predictable, high-volume production workloads. Overkill (and expensive) while you're experimenting.

For learning and prototyping, Standard or Global Standard is the right call. Give the deployment a clear name (you'll type this name in your code later, so make it something you'll recognize, like gpt-5-main).

Click deploy, wait a few seconds, and — that's it. You've deployed your model.

Step 4 — Test it in the Chat Playground

You don't need to write a line of code to confirm it works. In the left menu, open Playground → Chat playground, select your new deployment from the dropdown at the top, and type a prompt. You're now talking to your GPT-5 deployment. This is also the best place to experiment with the system prompt and parameters (temperature, max tokens) before you bake them into an app.

Try something that exercises the model, not just "hi" — for example: "Explain the difference between a model and a deployment in Azure AI Foundry to a junior developer, in three sentences." If you get a sensible answer, your endpoint is live and healthy.

The gotchas nobody mentions

A few things that will save you a support ticket:

Quota and region. New subscriptions start with modest token-per-minute quotas, and the newest models aren't available in every region. If deployment fails, it's almost always a quota or region issue — check the error message, it usually names the exact limit. You can request a quota increase from the portal.

Cost is per token, and it adds up quietly. The Playground is billed just like API calls. Keep an eye on the pricing on each model card, and set a budget alert on the resource group if you're on a personal subscription.

Deployment name vs. model name. In your code you reference the deployment name you chose, not the underlying model name. Mixing these up is the number-one cause of "deployment not found" errors.

What's next

Now that your model is deployed, the natural next step is calling it from your own code instead of the Playground — connecting to the endpoint from a Python project in VS Code, sending prompts programmatically, and building something real on top of it. That's exactly what I cover in my follow-up post.

Deploying your first model feels like a big deal the first time, and then it becomes second nature. The important thing is that you now own the full loop: discover a model, deploy it, and talk to it. Everything else you build on Azure AI is a variation on this foundation.

Happy building!