The GPT-5 model family, Azure AI Foundry architecture, Provisioned Throughput pricing, and a working technical guide for migrating off the direct OpenAI API — everything a Microsoft-first enterprise needs to plan an Azure OpenAI deployment.
Data verified July 2026 · sourced from Microsoft's official Azure OpenAI documentation
Azure OpenAI Service is Microsoft's enterprise-hosted version of OpenAI's models — GPT-5, GPT-4.1, the o-series reasoning models, and GPT-4o — deployed inside your own Azure subscription. Instead of calling OpenAI directly, you get the same models wrapped in Azure's identity layer (Microsoft Entra ID), private networking (VNet, Private Link), and compliance certifications (SOC 2, HIPAA, FedRAMP High). It's managed today through Azure AI Foundry, the unified portal that replaced the standalone Azure OpenAI Studio.
From the flagship GPT-5 to cost-optimized nano variants and the wider Azure AI Foundry model catalog.
Complex reasoning, multi-step agentic workflows, and highest-quality generation across enterprise workloads.
High-volume classification, summarization, and chat where latency and cost matter more than peak reasoning depth.
Large-document analysis, codebase-wide reasoning, and retrieval-heavy applications needing very long context.
Step-by-step chain-of-thought tasks — math, scientific analysis, and multi-stage planning that need deliberate reasoning.
Vision, audio, and text together in one call — document understanding, image Q&A, and voice-enabled assistants.
Llama, Mistral, DeepSeek, and other open-weight models served serverlessly alongside GPT models in the same Foundry project.
Pay-as-you-go pricing per 1 million input tokens — output tokens are billed separately and typically cost 4-8x more.
USD per 1 million input tokens
Azure OpenAI offers four ways to buy capacity — three pay-as-you-go tiers plus reserved Provisioned Throughput.
Requests route to whichever Azure region has capacity, giving the best throughput and lowest cost per token for workloads without strict data-residency needs.
Keeps processing within a geography (e.g. US or EU) for regulatory requirements, at a small premium over Global Standard.
Pins processing to one specific Azure region — the strictest data-residency option, typically at the highest per-token rate of the three pay-as-you-go tiers.
Dedicated, reserved model capacity starting around $2,448/month, with 15-50 PTU minimums depending on deployment type. Can cut effective per-token cost by up to 70% once volume is high and consistent enough to justify the commitment.
Keyless, identity-based authentication instead of long-lived API keys — the recommended pattern for production Azure workloads.
Built-in content filtering for hate, violence, self-harm, and sexual content categories, plus prompt-injection ("jailbreak") detection.
VNet integration and Azure Private Link keep model traffic off the public internet, meeting the same network controls as the rest of your Azure estate.
SOC 2 Type II, ISO 27001, HIPAA (with a signed BAA), and FedRAMP High in Azure Government — audited alongside your existing Azure compliance posture.
A unified hub-and-project structure for organizing models, agents, evaluations, and deployments — the successor to the standalone Azure OpenAI Studio.
The same Azure OpenAI models power Microsoft 365 Copilot and GitHub Copilot Enterprise, so custom apps and off-the-shelf Copilots share one governance model.
For most codebases this is a small, mechanical change — not a rewrite.
Create an Azure OpenAI resource inside an Azure AI Foundry project, then deploy a model — this generates a deployment name distinct from the underlying model name.
Resource creation, model deployment, quota (TPM) planning
Replace your OpenAI API key with an Azure OpenAI endpoint and key — or better, Microsoft Entra ID for keyless auth in production.
AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_API_KEY, or Entra ID token
Point your OpenAI SDK at your Azure endpoint using the v1 API path — the new v1 surface no longer requires an api-version parameter on every call.
base_url: https://YOUR-RESOURCE.openai.azure.com/openai/v1/
Azure OpenAI always requires the deployment name in requests — even when your code passes what looks like a model name, it must map to a deployment you created.
Deployment-to-model mapping, per-environment deployment names
Azure AI Foundry's portal generates working code snippets prefilled with your project's variables — run those to confirm parity before routing production traffic.
Side-by-side output comparison, latency and rate-limit testing
General availability with GPT-3.5 and early GPT-4 access, deployed through Azure OpenAI Studio.
Provisioned Throughput Units roll out across Global, Data Zone, and Regional tiers with lower minimum commitments.
Azure OpenAI Studio is absorbed into Azure AI Foundry — a single hub-and-project workspace spanning OpenAI models, the broader model catalog, and agent tooling.
GPT-5, GPT-5-mini, and GPT-5-nano land in Azure OpenAI alongside a simplified v1 API surface that drops the mandatory api-version parameter.
Custom assistants embedded in internal tools, extending the same Copilot experience employees already use in Microsoft 365.
Real-time call summarization, sentiment detection, and agent-assist copilots layered onto existing telephony and CRM systems.
GitHub Copilot Enterprise and custom coding assistants built on the same GPT-5 and o-series models available through Azure OpenAI.
Retrieval-augmented generation over contracts, policies, and knowledge bases using Azure AI Search alongside Azure OpenAI embeddings.
The full three-way comparison — model breadth, pricing model, and governance — plus an interactive model explorer.
Compare all three →Gemini 3.5, Agent Builder, and Google's newly rebranded enterprise AI platform, explained in the same depth.
Read the Gemini guide →We'll assess your workloads, right-size PTU vs. pay-as-you-go, and handle the migration from your existing OpenAI integration.