Study guide
6.1 Azure OpenAI Service — Models and Deployment
Azure AI Engineer Associate AI-102
Track your progress as you read
Finished sections get a check mark.
1 Introduction
About the Azure AI Engineer Associate (AI-102) Exam Study Strategies and Exam Tips for the AI-102
2 Domain 1: Plan and Manage an Azure AI Solution (20-25%)
1.1 Azure AI Services Overview and Selection 1.2 Resource Provisioning and Configuration 1.3 Security and Access Management 1.4 Responsible AI Principles and Implementation 1.5 Azure AI Solution Architecture Patterns
3 Content Safety and Moderation (within Plan and Manage, Domain 1)
2.1 Azure AI Content Safety Overview 2.2 Prompt Shields and Adversarial Attack Detection 2.3 Text and Image Moderation Implementation 2.4 Content Filtering in Azure OpenAI Service
4 Domain 4: Implement Computer Vision Solutions (10-15%)
3.1 Azure AI Vision — Image Analysis 4.0 3.2 Azure AI Face Service 3.3 Azure AI Custom Vision 3.4 Optical Character Recognition (OCR) 3.5 Video Analysis and the Retired Spatial Analysis
5 Domain 5: Implement Natural Language Processing Solutions (15-20%)
4.1 Azure AI Language — Text Analytics Features 4.2 Conversational Language Understanding (CLU) 4.3 Custom Text Classification and Custom NER 4.4 Azure AI Speech Services 4.5 Azure AI Translator Service 4.6 Question Answering and Conversational AI
6 Domain 6: Implement Knowledge Mining and Information Extraction Solutions (15-20%)
5.1 Azure AI Search — Fundamentals and Architecture 5.2 AI Enrichment with Skillsets 5.3 Azure AI Document Intelligence
7 Domain 2: Implement Generative AI Solutions (15-20%)
6.1 Azure OpenAI Service — Models and Deployment 6.2 Chat Completions API and Prompt Engineering 6.3 Embeddings and Vector Search for RAG 6.4 Fine-Tuning and Model Customization 6.5 Image Generation (gpt-image; DALL-E 3 Retired) 6.6 RAG After On Your Data, and AI Agents
8 Domain 3: Implement an Agentic Solution (5-10%)
3.1 Agentic Solutions: Foundry Agent Service and the Microsoft Agent Framework
9 Exam Review: Cross-Domain Topics and Advanced Practice
7.1 Service Comparison and Selection Guide 7.2 SDK and REST API Patterns 7.3 Responsible AI Across All Domains 7.4 End-to-End Solution Architectures 7.5 Exam Day Quick Reference
6.1 Azure OpenAI Service — Models and Deployment
Key Takeaways Explain This Topic AI
Azure OpenAI Service (now part of Microsoft Foundry Models) exposes OpenAI models such as GPT-4o, GPT-4.1, o-series reasoning models, gpt-image models, Whisper, and text-embedding-3 inside Azure with enterprise security.
A model is unusable until you create a deployment: a named endpoint binding one model version to a capacity quota expressed in thousands of tokens per minute (TPM).
Deployment types in 2026 are Standard, Global Standard, Data Zone Standard, Provisioned-Managed (PTU), Global/Data Zone Provisioned, and Batch — they differ on data-residency scope and on pay-per-token versus reserved-capacity billing.
Azure OpenAI models deploy into a Microsoft Foundry resource (kind AIServices), which reaches Azure OpenAI and other Foundry Tools through one endpoint and key, or into a standalone Azure OpenAI resource (kind OpenAI).
Tokens (~0.75 word each in English) drive both cost (priced per 1,000 input and output tokens separately) and the context-window and TPM limits you must size around.
Last updated: October 2026
Quick Answer: Azure OpenAI (now surfaced through Microsoft Foundry Models ) gives you GPT-4o/GPT-4.1 (multimodal text+vision), o-series reasoning models, gpt-image models (images; DALL-E 3 was retired March 4, 2026), Whisper (speech-to-text), and text-embedding-3 (vectors). Nothing works until you create a deployment that binds one model version to a TPM quota. Pick a deployment type — Standard, Global Standard, Data Zone, Provisioned-Managed (PTU), or Batch — based on data residency and traffic predictability.
The AI-102 Context
Generative AI (Domain 2) is 15-20% of AI-102 ( Designing and Implementing a Microsoft Azure AI Solution ), among the most heavily revised areas in 2026. The exam runs 100 minutes , passing at 700 / 1000 , costing about USD $165 . Microsoft retired the AI-102 exam on June 30, 2026; AI-103 replaced it.
Model Catalog You Must Recognize
- Model Type What it does Context window
- GPT-4o / GPT-4o mini Multimodal Text + image input, fast general chat 128K tokens
- GPT-4.1 / 4.1 mini Text + vision Long-context coding & instruction following up to 1M tokens
- o-series (o1, o3-mini) Reasoning Step-by-step reasoning; no temperature control 128K-200K
- gpt-image series Image generation Text-to-image N/A
- Whisper Audio Speech-to-text & translation 25 MB / file
- text-embedding-3-large Embeddings 3,072-dim vectors 8,191 tokens
- text-embedding-3-small Embeddings 1,536-dim vectors, cheaper 8,191 tokens
On the Exam: o-series reasoning models ignore temperature / top_p and use max_completion_tokens (reasoning tokens are billed but hidden). A question that sets temperature=0 on o1 is a distractor — the parameter is silently ignored.
Creating the Resource
Azure OpenAI models deploy into either a Microsoft Foundry resource ( --kind AIServices ), which also reaches Vision, Language, Speech, and the other Foundry Tools through one endpoint and key, or a standalone Azure OpenAI resource ( --kind OpenAI ), shown here. az cognitiveservices account create \
- -name my-openai-service \
- -resource-group rg-ai-prod \
- -kind OpenAI \
- -sku S0 \
- -location eastus2 \
- -yes
The key flag is --kind OpenAI for a standalone resource (use --kind AIServices for a Foundry resource). Either way, a GPT-4o call fails until you create a deployment for the model.
Deploying a Model
A deployment is a named alias mapping one model + version to a capacity quota. You call the deployment name , never the raw model name. az cognitiveservices account deployment create \
- -name my-openai-service \
- -resource-group rg-ai-prod \
- -deployment-name gpt4o-chat \
- -model-name gpt-4o \
- -model-version "2024-11-20" \
- -model-format OpenAI \
- -sku-name GlobalStandard \
- -sku-capacity 50 # 50K tokens-per-minute quota
- Deployment Types (2026)
- Type Data residency Billing Best for
- Standard Single Azure region you pick Pay-per-token Region-locked workloads
- Global Standard Routed to any global datacenter Pay-per-token, highest quota Default for dev/variable traffic
- Data Zone Standard Stays inside US or EU zone Pay-per-token EU/US data-residency rules
Provisioned-Managed (PTU) Region / Data Zone / Global Reserved capacity (hourly/monthly) High-volume production, stable latency
Batch Global ~50% cheaper, 24h SLA Large async offline jobs
Provisioned Throughput Units (PTU): you reserve a fixed block of capacity for guaranteed throughput and predictable latency. PTUs lower per-token cost at high volume but require a commitment. Choose PTU when traffic is predictable and high-volume ; choose Standard/Global for spiky or experimental loads.
Azure OpenAI vs. OpenAI Direct
- Feature Azure OpenAI OpenAI Direct
- Auth API key or Microsoft Entra ID / managed identity API key only
- Network VNet, private endpoints, IP firewall Internet only
- Content safety Built-in, configurable filters Limited
- Compliance SOC 2, HIPAA, ISO, FedRAMP Limited
- Data use Prompts/outputs not used to train models Not used (API)
- SLA 99.9% Standard / higher for PTU Best effort
- Region You choose the Azure region OpenAI-managed
On the Exam: "Why choose Azure OpenAI over OpenAI?" → enterprise security and compliance: managed identity, private endpoints, content filtering, and regional/data-residency control . Prefer managed identity over API keys for any production scenario the exam describes.
Tokens, Quota, and Cost
Tokens are the billing and limiting unit. In English, 1 token ≈ 0.75 words ≈ 4 characters , so 1,000 words ≈ 1,333 tokens. Total request tokens = input (prompt) + output (completion) , each priced separately per 1,000 tokens.
- Limit Meaning
- Context window Max input + output tokens per single request
- Max output tokens Cap on the completion length
- TPM (tokens/min) Throughput quota assigned to the deployment
- RPM (requests/min) Derived rate limit (roughly TPM/1000 × 6)
Worked example: A summarizer sends a 6,000-token document plus a 200-token instruction and expects a 600-token summary. Per call ≈ 6,800 tokens. A 30K-TPM Standard deployment supports ≈ 30,000 / 6,800 ≈ 4 calls/minute before throttling (HTTP 429). To scale, raise --sku-capacity (quota) or move to Global Standard / PTU. Common trap: developers blame the model for 429s when the real fix is increasing TPM quota or using Retry-After backoff.