Azure OpenAI / Azure AI Foundry — Library index

Canonical reference for Microsoft’s enterprise model-hosting path — the Azure OpenAI Service (OpenAI’s GPT / o-series / image / audio / video / embedding models sold by and operated by Azure) plus the broader Azure AI Foundry platform (formerly Azure AI Studio, now folded under the “Microsoft Foundry” brand) that wraps it: a 1,800-plus-model catalog, Models-as-a-Service serverless endpoints for Llama / Mistral / Cohere / DeepSeek / xAI, the Foundry Agent Service, prompt flow, evaluations, and Azure AI Search for RAG. House style and source URLs match the rest of the Compute library; this sub-library is the Azure-specific counterpart to AWSBedrock/ (Amazon’s hyperscaler model gateway) and GoogleGemini/’s Vertex notes (Google’s). Where Bedrock’s headline is “third-party models behind one AWS IAM-controlled API,” Azure’s headline is “OpenAI’s frontier models first, behind one Entra-ID-controlled API, with the strongest enterprise compliance / data-residency story of the three.”

Azure occupies a distinctive seat in the AI stack. Through the Microsoft–OpenAI partnership it has historically been the only enterprise cloud to host OpenAI’s frontier line (GPT-5.x, o-series, Sora, gpt-image) under a hyperscaler’s SLA, billing, and compliance umbrella. That changed in form but not in practice on 2026-04-27, when Microsoft and OpenAI restructured the deal: OpenAI may now sell across any cloud and the exclusivity clause is gone, but Microsoft’s model license runs through 2032 and it still gets new OpenAI models first — GPT-5.5 shipped in Foundry the day after the announcement (azure.microsoft.com/blog 2026-04). For an enterprise that lives in Entra ID, Microsoft 365, and Azure compliance boundaries, Foundry is the path of least resistance to OpenAI models in production, and increasingly to the rest of the open and partner model ecosystem too.

This sub-library lives inside Math-and-Compute/Compute/ as a sibling of Claude/, OpenAI/, GoogleGemini/, AWSBedrock/, Vercel/, and CloudflareWorkers/. Notes are version-pinned to 2026-05-30. Azure’s surface moves fast and renames fast — “Azure AI Studio” became “Azure AI Foundry” (late 2024), and through 2025–2026 the docs migrated again to the “Microsoft Foundry” brand with the OpenAI piece relabeled “Azure OpenAI in Microsoft Foundry Models.” Both portals (the new Foundry portal and “Foundry classic”) coexist. Verify model IDs, deployment-type availability, pricing, and regional coverage against learn.microsoft.com/azure/foundry/ before quoting.

Azure OpenAI Service vs Azure AI Foundry — the two layers

These are easy to conflate. The distinction drives almost every architecture decision in this library:

• Azure OpenAI Service (the model layer) — the resource type (Microsoft.CognitiveServices/accounts of kind OpenAI/AIServices) that exposes OpenAI’s models on Azure infrastructure. This is where GPT-5.x, o-series, embeddings, Whisper, gpt-image, Sora, and the audio/realtime models live. You call it with the OpenAI-compatible SDK pointed at your Azure endpoint, authenticated by Entra ID or key. Billed through your Azure subscription, covered by an Azure SLA, supported by Microsoft. These are the “Foundry Models sold by Azure” (a.k.a. “Direct from Azure” / “Azure Direct” models) — hosted and operated by Azure.
• Azure AI Foundry (the platform layer, formerly Azure AI Studio) — the unified studio + SDK + API on top: a model catalog of 1,800-plus models, Models-as-a-Service serverless endpoints and managed-compute for partner/open models (Llama, Mistral, Cohere, DeepSeek, xAI, Phi), prompt flow orchestration, an evaluation harness, the Foundry Agent Service, and integrations with Azure AI Search (RAG), Bing, SharePoint, and Microsoft Fabric. Azure OpenAI is the most important tenant of this platform, not a separate product. “Foundry Models from partners and community” are catalog models hosted under partner terms rather than sold directly by Azure.

The single deep note in this library covers both layers end-to-end; this index explains where each piece lives.

See also

• _index — the model-provider deep dive for the OpenAI family (the canonical reference for GPT-5.x / o-series capabilities, API shape, and pricing on OpenAI’s own platform; Azure OpenAI re-hosts these models but is not the canonical model reference for them — much like AWSBedrock/ re-hosts Claude but defers to Claude/)
• _index — Anthropic ships Claude on Foundry too (“Microsoft Foundry” is one of Anthropic’s listed distribution channels), but Claude/ is the canonical Claude reference
• _index — the AWS counterpart: the direct structural sibling to this library. Read both side by side when choosing a hyperscaler model gateway
• vertex-ai-enterprise-gemini — the Google counterpart (Vertex AI), the third leg of the hyperscaler-model-hosting comparison
• llm-landscape — vendor-neutral catalog of the whole model field; reconcile any model-version claim here against it
• cloud-provider-service-mapping — how Azure’s AI services map to AWS and GCP equivalents
• rag-embeddings-vector-search — the from-scratch RAG path that Azure AI Search + Foundry’s grounding tools provide as a managed alternative
• prompt-engineering-agent-systems — vendor-neutral discipline overview that the Foundry Agent Service operationalizes
• auth-authz — Microsoft Entra ID is how every Azure OpenAI call should be authorized in production
• observability-stack — Azure Monitor / Application Insights is the native observability target for Foundry workloads
• model-serving-infrastructure — what the managed-compute serving layer is doing under the hood for open-weight models

Library contents

Foundations

This library currently ships a single dense note rather than Bedrock’s seven-note split. That is deliberate: Azure’s value-added stack (Agent Service, AI Search RAG, Content Safety) overlaps heavily with services documented vendor-neutrally elsewhere in Compute, so the Azure-specific facts collapse cleanly into one lookup. If the Agent Service or Foundry evaluations grow enough Azure-specific surface to warrant their own notes, split them out as agent-service.md and foundry-evaluations.md then — peers to Bedrock’s bedrock-agents.md.

How to use this library

1. Picking a model for a new Azure project — start in azure-openai-and-ai-foundry §“Model catalog” for the lineup, then §“Deployment types” for the call-shape / quota decision, then §“PTU economics” if volume is high enough to justify reserved capacity.
2. Choosing between Azure, Bedrock, and Vertex — go straight to azure-openai-and-ai-foundry §“Honest comparison,” cross-read with bedrock-models-and-providers and vertex-ai-enterprise-gemini.
3. Meeting a data-residency / compliance requirement — §“Regional availability & data residency” + the Data Zone deployment types are the enterprise selling point; §“Identity & networking” covers Private Link / VNet isolation.
4. Building a RAG application on Azure — §“Foundry platform” (Azure AI Search grounding) is the managed path; rag-embeddings-vector-search is the from-scratch path.
5. Building an agent — §“Foundry Agent Service” is the managed orchestrator; tool use + the Responses API are the lower-level building blocks.
6. Safety / Responsible AI — §“Content filtering & Responsible AI” covers what the managed safety layer (Prompt Shields, groundedness, Protected Material) does.

Conventions used in this library

• Model IDs are quoted verbatim as Azure deployment model names (e.g. gpt-5.5, gpt-5.2, o4-mini, text-embedding-3-large, gpt-image-1.5, sora-2). Azure model names usually match OpenAI’s but the deployment you call is a name you choose at deploy time, decoupled from the model ID.
• Deployment-type SKU codes are quoted verbatim (GlobalStandard, GlobalProvisionedManaged, DataZoneStandard, DataZoneProvisionedManaged, ProvisionedManaged, GlobalBatch, DataZoneBatch, DeveloperTier).
• Pricing is in USD per million tokens (MTok) on the relevant deployment type as of 2026-05-30. Where multiple \$ symbols appear on the same line they are backslash-escaped to avoid Obsidian MathJax interpretation.
• “Per learn.microsoft.com 2026-05” tags a claim sourced from official Microsoft Foundry documentation as of the file’s created date.
• “Foundry” without qualification means the current Microsoft Foundry brand; “Foundry classic” denotes the legacy portal/docs that still coexist. “Azure AI Studio” is the deprecated former name of the platform layer.

Sources read while building this library (2026-05-30 pass)

• learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure (model catalog, capabilities, context windows)
• learn.microsoft.com/azure/foundry/foundry-models/concepts/deployment-types (all eight deployment types, SKU codes, data-residency scopes)
• learn.microsoft.com/azure/foundry/openai/concepts/provisioned-throughput (PTU concept, quota vs capacity, hourly vs reservation billing, spillover)
• learn.microsoft.com/azure/foundry/agents/overview + Agent Service tool-catalog / MCP pages
• learn.microsoft.com/azure/ai-services/content-safety/overview + /whats-new (severity levels, Prompt Shields, groundedness, Protected Material)
• learn.microsoft.com/azure/ai-services/openai/how-to/managed-identity + network / Private Link pages
• azure.microsoft.com/pricing/details/azure-openai/ (pay-as-you-go and PTU pricing structure)
• azure.microsoft.com/blog GPT-5 / GPT-5.2 / GPT-5.5 in Foundry announcements; the 2026-04-27 partnership restructuring coverage

Where official docs are silent but behavior is observable (exact PTU-to-TPM ratios per model, real-time capacity volatility, undocumented header behavior), entries in the deep note are labelled “Undocumented as of 2026-05-30 — observed / aggregator-sourced” so a future reader knows to verify against learn.microsoft.com.