Azure OpenAI: a practical guide for SMEs

What is Azure OpenAI?

Azure OpenAI is Microsoft's managed service for using OpenAI's language models – the GPT family and the o reasoning models – via API, without running infrastructure yourself. You send a prompt, get a response and pay only for the tokens you use. In early 2026 the platform was renamed “Microsoft Foundry”; it's still sold and billed through Azure.

The difference from the direct OpenAI API: the same models, but inside Microsoft's enterprise environment. Authentication runs through Entra ID (formerly Azure AD) instead of a shared API key, you get SLAs, private networking and – the point of this guide – the option to keep processing inside the EU.

For companies already using Microsoft 365 and Azure, that's appealing because everything lives in the same world: same billing, same access management, the same data-processing agreement your legal team already signed.

Why Azure OpenAI for SMEs? The GDPR argument

The biggest hurdle for AI in German SMEs is rarely the technology – it's data protection. “Are we even allowed to send customer data to a US model?” stalls almost every project.

Azure answers this with the EU data zone. Create your deployment as “Data Zone Standard” in an EU region and prompts and responses are processed exclusively within the EU Data Boundary. Microsoft does not use your inputs and outputs to train the models, and the Microsoft DPA (the Article 28 GDPR data-processing agreement) applies to Azure customers automatically.

Important: a “Global” deployment does NOT. There, processing can happen in any region worldwide – only data at rest stays in your geography. For real EU data residency you must deliberately choose a Data Zone or single-region deployment. That switch is the heart of the matter.

How processing stays in the EU

CurrentModels with an EU data zone

You can run these OpenAI models on Azure with an EU data zone (Data Zone Standard in an EU region). Prices are on-demand list prices per 1M tokens in US dollars (Global Standard) – the EU data zone costs the same.

The cheaper mini/nano models (GPT-5.4 mini/nano) are currently only available as Global or US Data Zone deployments, not in the EU data zone – the cheapest EU-resident model right now is GPT-5. And Azure's o reasoning models (o3, o4-mini) retire at the end of 2026; check the model lifecycle before you commit.

As of 7 July 2026. Prices and availability change – the Azure pricing page is authoritative. Official Azure pricing ↗

CurrentPrice range: cost per 1M tokens

What does Azure OpenAI cost?

Azure bills on-demand by default: you pay per token, separately for input and output. No base fee, no idle cost – send nothing, pay nothing.

The pleasant part for the EU: a Data Zone deployment costs the same per-token price as Global Standard. Unlike some competitors, you pay no surcharge for EU processing – only the strictest variant (processing in exactly ONE region) runs about 5–10% higher.

For predictable, high load Azure offers “Provisioned Throughput” (reserved capacity in PTUs). For most SMEs that's unnecessary at first: start on-demand, measure your real usage, then optimize. Summarizing two A4 pages with GPT-5 costs fractions of a cent – costs only become noticeable at volume.

CurrentCost calculator: what does it cost per month?

Step by step: your first call on Azure OpenAI

Here's how to go from zero to your first response – in about 15 minutes. Azure OpenAI is self-service in 2026; standard use no longer needs an access request.

1. 1
   Sign in with an Azure subscription

   Sign in with an Azure subscription in the portal or in Microsoft Foundry (ai.azure.com). Standard use needs no approval form.

2. 2
   Create a resource in an EU region

   Create an Azure OpenAI / Foundry resource and pick an EU region such as Sweden Central, Germany West Central or France Central – so the resource sits inside the EU Data Boundary.

3. 3
   Deploy the model as “Data Zone Standard”

   In the deployment dialog, choose a model (e.g. gpt-5) and the “Data Zone Standard” deployment type. Give the deployment a name – you pass that as model in your code, not the base model name.

4. 4
   Get the endpoint + key (or Entra ID)

   Copy the endpoint and one of the two keys from “Keys and Endpoint”. Cleaner for production: keyless via Entra ID with the “Cognitive Services User” role and DefaultAzureCredential.

5. 5
   Install the SDK and make the first call

   pip install openai, set the key as an environment variable and call the AzureOpenAI client – see the code below. The response comes back structured.

import os
from openai import AzureOpenAI

client = AzureOpenAI(
    api_version="2024-10-21",
    azure_endpoint="https://DEIN-RESOURCE.openai.azure.com",
    api_key=os.environ["AZURE_OPENAI_API_KEY"],
)

# "model" = dein Deployment-Name (Data Zone Standard) – nicht der Basis-Modellname.
response = client.chat.completions.create(
    model="gpt-5-eu",
    messages=[
        {"role": "user", "content": "Erklär in einem Satz, was Azure OpenAI ist."}
    ],
)

print(response.choices[0].message.content)

The model value is your deployment name (Data Zone Standard in an EU region) – not the base model name. That keeps processing in the EU.

Azure OpenAI, direct API or Bedrock? A clear take

Azure isn't always the best choice. The quick decision guide:

Since April 2026 the Microsoft–OpenAI license is no longer exclusive – GPT models are on Bedrock too. “We need GPT” no longer ties you to Azure automatically. What matters is where you already are.

GDPR checklist for Azure OpenAI

To use Azure OpenAI in a data-protection-compliant way, watch these points:

• ✓Choose the “Data Zone Standard” deployment type (or a single-region deployment) in an EU region – NOT “Global Standard”. That's the only way processing is guaranteed to stay in the EU.
• ✓Data processing: the Microsoft DPA (aka.ms/DPA) applies to Azure automatically and includes the Standard Contractual Clauses – with an Enterprise Agreement / Microsoft Customer Agreement, no extra step.
• ✓No training: Microsoft does not use your prompts and responses to train the models; Azure OpenAI runs separately from OpenAI's own services.
• ✓Decide on abuse monitoring deliberately: by default inputs and outputs can be stored for up to 30 days. For zero storage, apply for “Modified Abuse Monitoring” / Zero Data Retention (usually EA/MCA).
• ✓Access via Entra ID + roles instead of a shared API key; for sensitive workloads use Private Link / VNet.
• ✓Minimize data in the prompt and add the AI processing to your records of processing activities (and a DPIA if needed).

CurrentLatest Azure OpenAI news

What's happened recently around Azure OpenAI / Microsoft Foundry:

• 24 April 2026GPT-5.5 in Microsoft Foundry — frontier model on an enterprise platform (~1.05M-token context)Source: Microsoft ↗
• 17 March 2026GPT-5.4 mini & nano — low-latency, low-cost models in Microsoft FoundrySource: Microsoft ↗
• 11 December 2025GPT-5.2 in Microsoft Foundry — the new enterprise standardSource: Microsoft ↗

Frequently asked questions about Azure OpenAI