Azure AI Foundry

Separate the platform, model originator, and deployment

The current official name is Microsoft Foundry; Azure AI Foundry remains a common former name. It is Azure's multi-provider catalog and deployment platform, not the originator of every listed model. Azure OpenAI, Microsoft models, and partner models can use different terms, regions, interfaces, and content-safety configuration. This site's /v1/chat/completions is only a gateway protocol for configured channels.

Requests normally use a deployment name

A catalog model is first deployed as an Azure resource with a caller-defined deployment name. Request model normally uses that deployment name, not the model ID shown in the catalog. The deployment also binds version, capacity type, filters, and rate limits. For 404, check resource endpoint, deployment name, and API version before trying another model alias.

Prefer Microsoft Entra ID in production

Foundry supports API keys and Microsoft Entra ID. A key is convenient for a quick test, but production should prefer keyless Entra authentication with least-privilege roles such as Cognitive Services User; the token scope must match the target service. Store and rotate any API key server-side. This site's token, an Azure key, and an Entra access token are not interchangeable.

Interfaces and SDKs are migrating

Azure OpenAI endpoints, Foundry Models endpoints, OpenAI/v1, and provider-specific paths do not have identical capabilities. Microsoft has deprecated the Azure AI Inference beta SDK and plans retirement on 2026-08-26, recommending the GA OpenAI/v1 API with stable OpenAI SDKs. Migrate base URL, deployment name, API version, response fields, tools, and stream events together.

Billing follows deployment type

Browsing Foundry does not make inference free. Language, vision, and audio can be charged by tokens or other per-model meters. Provisioned throughput, managed compute, and fine-tuned models can also incur reserved capacity, training, and continuous hosting cost. Configure budgets, Cost Management, Azure Monitor, and usage alerts for every deployment, and confirm that experimental hosting resources are removed when finished.

429, errors, and safety

For 401/403, inspect Entra scope, role, key, and model entitlement; for 404, inspect endpoint and deployment name. A 429 can mean TPM/RPM, saturated PTU, or shared capacity. Honor retry-after-ms, smooth concurrency, and use bounded backoff before changing quota or deployment type. Preserve request ID, effective model version, region, and filter results. Platform content safety does not replace application authorization, audit, or privacy controls.