Azure OpenAI gives you OpenAI's models behind your organisation's own Azure subscription, identity and network controls, at OpenAI's usual pace of model updates but with Microsoft's billing and compliance paperwork. Pick it when your organisation already runs on Azure; pick a direct provider key when it doesn't and you just want the model.
Other good fits
A team that needs a specific data-residency option (Global, Data Zone, or one of ~27 regions) for its model traffic.
+ Pros
Read pros
Azure OpenAI puts OpenAI's models behind an Azure resource: you deploy a model under your subscription and call it with Azure's authentication (Entra ID or an API key) instead of an OpenAI account. That means model usage shows up on your existing Azure bill and can sit behind your organisation's existing network and identity rules.
Pick it when…
Your organisation already manages access through Entra ID and wants model calls to use the same identity system.
You need to choose where requests are processed: Global, a Data Zone (EU or US), or one of roughly 27 regional deployments.
Provisioned throughput (a reserved capacity you pay for monthly or annually) matters more to you than pay-as-you-go pricing.
− Cons
Read cons
Not every model or feature is available in every region, so plan around your organisation's actual regional and compliance needs before assuming a model is available. store=False avoids Azure's default 30-day stored-response retention, but it isn't a blanket promise that no record of the request exists anywhere in the pipeline; check the current data-privacy documentation for what still applies.
Pick something else when…
You don't already use Azure. Call the model directly through the Claude API or OpenAI's own API and skip the resource-and-deployment setup.
You want one key across many providers instead of managing Azure deployments per model.OpenRouter or Requesty route to multiple providers from one endpoint.
You want EU-hosted inference without Azure's regional deployment matrix.EUrouter is built around European infrastructure specifically.
Avoid if
You need the newest OpenAI model the day it ships; Azure's catalogue and regional availability can lag the direct API.
Connection
Create an Azure OpenAI resource, deploy a specific model under a deployment name, then call https://YOUR-RESOURCE-NAME.openai.azure.com/openai/v1/ with either an API key or Entra ID authentication. The deployment name, not the model's public name, is what you send in each request, and it only exists in your subscription.
Per input/output token, billed like any Azure consumption
Provisioned Throughput Units
Reserved monthly or annual capacity, billed whether or not you use it
Batch API
Same models at roughly half the Global Standard rate, results within 24 hours
Azure doesn't publish a single flat per-model rate on the page; exact numbers depend on the model, deployment type and your Microsoft agreement, so check the pricing calculator for your specific model before committing budget. The ongoing effort is the same as any Azure resource: managing deployments, quotas, access roles and regional choices, on top of reviewing what the model actually returns.
Data handling
Azure documents three processing boundaries: Global, Data Zone (EU or US) and Regional, each with different residency guarantees. store=False opts out of the default 30-day stored-response retention Azure otherwise applies. Confirm which boundary your deployment uses; the resource's address alone doesn't tell you.
First request
Documentation-based suggestion — not run here.
Deploy a small, inexpensive model and send one short, limited request: a two-sentence reminder generated from a short made-up note. Set store=False so the response isn't kept in Azure's stateful history, and read the actual response object rather than assuming it matches the plain chat-completions format.
Done when you've matched the response to the input, confirmed no invented detail was added, and checked the usage field against your deployment's quota.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on Microsoft's Azure OpenAI quickstart, pricing page and data-privacy documentation, read on 2026-09-23. No Azure resource was deployed or called for this review; exact per-model token rates were not published on the fetched pricing page.
Call OpenAI text, audio, and vision models from applications in Azure.
Choose it when Azure integration and managed regional deployment matter.
Batch processing and provisioned throughput support different operating patterns.