Guides · ai · Verification-gated
Run OpenAI models (and other "models sold by Azure") inside your own Azure subscription. You need an Azure account with a subscription, an Azure OpenAI or Foundry resource, and a model deployment. Most models need no application any more; quota, a few limited-access models and abuse-monitoring opt-outs are the remaining gates.
Researched · not yet run liveLast researched 25 October 2026 · Microsoft · provider site
A Microsoft or GitHub account signs up for Azure; that creates an Entra tenant and a billing account with a subscription. Inside the subscription you create a resource group, then an Azure OpenAI (or Foundry) resource in a region, then one or more model deployments. Keys belong to the resource; quota belongs to the subscription.
Microsoft recommends Microsoft Entra ID (for example a managed identity) instead of API keys.
Steps marked “human step” are ones only you can do: verapli never signs in, submits, pays or accepts terms for you.
Sign up with a Microsoft or GitHub account, verify a phone number and a non-prepaid card. Sign-up does not charge you; the free credit lasts up to 30 days.
Where: https://azure.microsoft.com/en-us/pricing/purchase-options/azure-account
Free accounts must move to pay-as-you-go within 30 days or when the credit runs out, or services are disabled. Free trials and student subscriptions also get no batch quota, so upgrade before real use.
Where: https://portal.azure.com/
In the Azure portal choose Create a resource > Azure OpenAI, then pick subscription, resource group, region, a resource name and the Standard pricing tier, and choose network access (all networks, selected networks or private endpoint only). You need RBAC permission to create the resource. The Foundry portal may offer to upgrade it to a Foundry resource.
Where: https://portal.azure.com/#create/Microsoft.CognitiveServicesOpenAI
In the Foundry portal open your resource > Deployments (or Models + endpoints) > Deploy model > Deploy base model, choose the model, a deployment name and a deployment type. Your code calls the deployment name.
Where: https://ai.azure.com/
On the resource's Keys and Endpoint page copy the endpoint and KEY1 (KEY2 is a spare for rotation). Microsoft recommends Microsoft Entra ID with managed identities instead of keys; if you use keys, keep them in a vault and rotate them.
verapli captures AZURE_OPENAI_API_KEY here by plain code: encrypted in your browser, never shown to AI.
Where: https://portal.azure.com/
Each subscription has a quota tier (Free Tier, Tiers 1-6) that rises automatically with use and payment history; enterprise agreements start higher. To get more now, submit the quota request form. Approved requests add quota without changing your tier.
Where: https://aka.ms/oai/stuquotarequest
A few models need registration, for example computer-use-preview (aka.ms/oai/cuaaccess) and xAI grok-4 / grok-code-fast-1 (aka.ms/xai/grok-4). Access is granted against Microsoft's eligibility criteria; approval for one limited model does not cover another.
Where: https://aka.ms/oai/cuaaccess
By default, flagged prompts and outputs can be stored and reviewed (by AI and, if needed, by authorised Microsoft staff) for abuse. Customers managed by a Microsoft account team or in an eligible programme can apply to turn off that storage and human review, or to switch guardrails partly or fully off. After approval, the resource's JSON view shows ContentLogging false.
Where: https://learn.microsoft.com/en-us/azure/ai-foundry/responsible-ai/openai/limited-access
Plain-English summary, not legal advice. Each card links its source.
Microsoft detects patterns such as repeated harmful content or jailbreak attempts. Confirmed abuse is notified by email; except for severe or recurring abuse you get a chance to fix it, but failure to fix, or severe or recurring abuse, can mean suspension or termination of access. [7]
Apps must disclose AI-generated output and voices, control misuse and offer abuse reporting. Banned uses include erotic or romantic chatbots, personas of real people without consent, impersonation, emotion inference and social scoring. Violators may lose access at Microsoft's discretion. [8]
Free credit lasts 30 days; if the account is not moved to pay-as-you-go, it and its services are disabled. [9]
Standard deployments bill per input and output token; provisioned throughput can be reserved monthly or yearly. Batch is 50% off Global Standard but returns within 24 hours. [11]
Prompts, completions, embeddings and training data are not shared with OpenAI and not used to train foundation models without your permission. Global and Data Zone deployments may process data outside your resource's region. [6]
Microsoft does not recommend preview models in production; preview deployments are upgraded to later versions. [2]
| Env var | Secret | Format | What it is |
|---|---|---|---|
AZURE_OPENAI_API_KEY | secret | opaque key string from Keys and Endpoint | Resource key (KEY1 or KEY2) sent in the api-key header. |
AZURE_OPENAI_ENDPOINT | not secret | https://<resource-name>.openai.azure.com/ | Resource endpoint base URL. |
Source [11]
The same entry as JSON: /api/v1/entries/microsoft/azure-openai · requirements /api/v1/requirements/microsoft/azure-openai · terms /api/v1/terms/microsoft/azure-openai.
claude mcp add --transport http verapli https://verapli.com/mcp
> "What do I need for Azure OpenAI in Microsoft Foundry Models, and how long will it take?"
→ tools/call requirements { "provider_id": "microsoft/azure-openai" }Provider rules change. This page is generated from verapli registry version 2026.1025.316; spotted something out of date? Tell us. Microsoft is not affiliated with verapli.