Azure OpenAI proxy: track and cap Azure OpenAI spend per feature with accurate deployment pricing
Route Azure OpenAI deployments through an OpenAI-compatible gateway. Use the standard OpenAI SDK, map deployment names to the real model for accurate cost, and enforce budgets before the call.
Azure OpenAI serves OpenAI models from your own Azure resource. Cognocient AI sits in front of it, so your application uses the plain OpenAI SDK and gets the same attribution, budgets and kill switch as every other provider.
Setup
In Settings → AI Providers → Azure OpenAI enter:
| Field | Example | Notes |
|---|---|---|
| API key | 3f9c... | Key 1 or Key 2 from the Azure portal |
| Endpoint URL | https://my-resource.openai.azure.com | The resource endpoint, without a path |
| Deployment name | prod-chat | The deployment calls are sent to |
| Underlying model | gpt-4o | Optional, but needed for correct cost if the deployment name isn't the model name |
Then call the gateway with the regular OpenAI client:
The gateway builds https://<resource>/openai/deployments/<deployment>/chat/completions?api-version=2024-10-21 and sends your key in Azure's api-key header. You don't use the AzureOpenAI client class or set an API version yourself.
Deployment names and accurate cost
Azure decides which model answers from the deployment name in the URL. The model field your app sends is not authoritative, and Azure ignores it for routing. If your deployment is called prod-chat, Cognocient AI can't tell from the request whether that is gpt-4o or gpt-4o-mini, a 16× price difference.
Fill in Underlying model once and every call is priced at that model's rate regardless of what the app sends. Leave it blank only if the deployment name is exactly the model name.
Routing between Azure and OpenAI
When an Azure key is configured, model names containing gpt-, o1 or o3 go to Azure; other names fall back to OpenAI. One Azure key maps to one deployment. To spread load across several deployments or regions, add more keys on Growth and above and use multi-key balancing, or put Azure and OpenAI behind one model name with load balancing.
Azure data residency
Prompts pass through Cognocient AI on their way to your Azure resource. Read Security & privacy for what is stored (metadata, not prompt content, by default) before routing regulated workloads.
Embeddings
/v1/embeddings works with Azure deployments of embedding models, priced per input token.
Related articles
Supported Providers
OpenAI, Anthropic, Gemini, Mistral, Groq, Together, Azure, AWS Bedrock, Vertex AI, and custom OpenAI-compatible endpoints.
OpenAI
OpenAI SDK through the gateway: chat, Responses, embeddings, images, audio and batches with per-feature cost and budgets.
Anthropic (Claude)
Claude through the OpenAI SDK (translated) or the native Anthropic SDK (passthrough), with prompt-cache-aware cost.