Azure OpenAI
Listed at https://azure.microsoft.com/en-us/products/ai-services/openai-service
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | GPT-4o: $2.50/M, GPT-4o-mini: $0.15/M, o1: $15/M, o3: $10/M | per million tokens |
| Output | GPT-4o: $10/M, GPT-4o-mini: $0.60/M, o1: $60/M, o3: $40/M | per million tokens |
🤖 Supported Models (8)
✨ Pros
- ✓Azure-hosted OpenAI models, data stays within Azure cloud
- ✓21Vianet provides compliant China access
- ✓Enterprise security/compliance (HIPAA/ISO/SOC2)
- ✓Provisioned Throughput for reserved capacity
- ✓Deep Azure ecosystem integration
- ✓Data not used for training (unlike direct OpenAI)
⚠️ Cons
- ×Pricing same as OpenAI official, no advantage
- ×International version not accessible in China; 21Vianet has fewer models
- ×Complex setup (requires Azure OpenAI application)
- ×Long approval process (3-5 business days average)
- ×Provisioned Throughput has minimum commitment
🎯 Best For
Enterprise customers needing Microsoft compliance; compliant OpenAI access in China; existing Azure customers
💰 Pricing & Plans
| Model | Input ($/M tokens) | Output ($/M tokens) | Best For |
|---|---|---|---|
| GPT-4o | $2.50 | $10.00 | Multimodal (text+image) |
| GPT-4o mini | $0.15 | $0.60 | High-volume, lightweight tasks |
| o1 | $15.00 | $60.00 | Advanced reasoning, long chain-of-thought |
| o3 | $10.00 | $40.00 | Structured reasoning, math/code |
| Provisioned Throughput (PTU) | Reserved hourly | — | Guaranteed capacity, low-latency SLAs |
| Pay-as-you-go | $0 | $0 | No commit; same per-token price as OpenAI direct |
🔧 API & Developer Experience
- •API Compatibility: 100% wire-compatible with the OpenAI /v1/chat/completions, /v1/embeddings, and /v1/images/generations REST surface; existing OpenAI SDKs work by swapping the base URL and adding an api-key header.
- •Endpoint Base URL: Per-deployment URL of the form https://{resource}.openai.azure.com/openai/deployments/{deployment}/ — your 'deployment' name acts as the model alias and lets you swap underlying models without code changes.
- •Model Availability: A subset of OpenAI's catalog — GPT-4o, GPT-4o-mini, o1, o3, GPT-4 Turbo, GPT-3.5 Turbo, DALL-E 3, Whisper, plus enterprise-only preview models (some only via Azure). Region- and quota-gated.
- •Auth & Identity: Azure AD (Entra ID) tokens or API keys; managed identities let apps running on Azure resources (App Service, Functions, AKS) call the API without storing secrets.
- •Content Safety: Azure AI Content Safety filter is layered on every prompt and completion by default — severity thresholds (Safe / Low / Medium / High) and category-level toggles are configurable per deployment.
- •Provisioning Model: Provisioned Throughput Units (PTU) reserve capacity for predictable low-latency SLAs; Pay-as-you-go (PTU-free) is identical per-token pricing to OpenAI direct, useful for spiky workloads.
- •Observability & Monitoring: Native Azure Monitor metrics, logs in Log Analytics, request/response traces in Application Insights; exportable to Power BI / Grafana via OpenTelemetry-compatible endpoints.
🏢 Enterprise Compliance & Data Residency
Azure OpenAI is the only mainstream OpenAI-compatible service that ships with the full Azure enterprise control plane built in. Prompts and completions live inside the customer's Azure tenant, are never used to train foundation models, and can be deleted on demand. It inherits Azure's compliance portfolio: SOC 2 Type II, ISO 27001 / 27018 / 27701, HIPAA BAA, FedRAMP High, and EU Data Boundary. Teams also get configurable Content Safety filters, Private Endpoint / VNet injection, customer-managed keys via Key Vault, and audit logs feeding Microsoft Purview. For regulated workloads, this is the single biggest reason teams standardize on Azure over the OpenAI direct API or third-party aggregators.
🌐 Regional Availability & Latency
Azure OpenAI has two distinct offerings for China. The global Azure OpenAI service (api.openai.azure.com across 60+ regions) is not reachable from mainland China due to cross-border data restrictions; it remains the right choice for international deployments via Hong Kong, Singapore, or Japan West. For Mainland-resident traffic, Microsoft 21Vianet operates a separate, ICP-licensed Azure China instance — access requires a Chinese Azure subscription (申请 Azure China account via 21Vianet), a separate approval process (typically 3–5 business days), and a smaller model catalog (some frontier models such as GPT-5 are not yet available on Azure China at launch). Latency from the 21Vianet instance to Chinese clients is the lowest among Western LLM providers — often under 80ms intra-region and 100–150ms cross-province — because traffic stays inside Mainland Chinese networks and never crosses the GFW. Combined with full ICP and MLPS 2.0 compliance, this is the standard answer for Chinese enterprises that need an OpenAI-grade model on compliant infrastructure.