Microsoft Foundry in Practice

Foundry vs. Direct API: The Current Feature Gap

The models are the same; the API surface is not. Before committing a workload to Claude on Foundry, check whether it depends on something that only exists on Anthropic's first-party API.

Claude 3P 101 · Updated July 2026 · Unofficial guide

Claude in Microsoft Foundry speaks the standard Messages API — same request shape, same usage object in responses, same model IDs (claude-opus-4-8, claude-sonnet-5, and so on, with no prefix). For a straightforward chat or tool-use workload, code written against the first-party API ports with little more than a client swap to AnthropicFoundry. The differences appear at the edges: whole API endpoints that do not exist on Foundry, features that depend on which hosting option you deployed, and operational conveniences that quietly are not there. As of July 2026, here is the map.

Not available on Foundry at all

Anthropic's Foundry documentation lists these as unsupported, full stop:

One operational gap deserves its own line: Foundry does not return Anthropic's anthropic-ratelimit-* response headers. Client-side limiters that read them must fall back to Azure-side monitoring and exponential backoff on 429s.

Available, but beta — and hosting-option dependent

PDF input, adaptive and extended thinking, prompt caching, citations, and 1M context are generally available on Foundry. Several other advanced features exist on Foundry in beta: structured outputs, token counting, web search, web fetch, code execution, tool search, the bash/text-editor/memory tools, Agent Skills, programmatic tool calling, the MCP connector, and the Files API (beta here, in contrast to Batches, which is absent). The caveats that come with the label are covered in beta feature caveats on Foundry.

The subtler axis is the hosting option chosen at deployment time — "Hosted on Azure" (recommended default, GA models) versus "Hosted on Anthropic infrastructure." Several features are only available on Anthropic-hosted deployments: structured outputs, the server-side tools (web search, web fetch, code execution, tool search), the MCP connector, Agent Skills, programmatic tool calling, and the Files API. Sending such a request to an Azure-hosted deployment returns 400 Bad Request by design. Note the trade-off is not free: on Anthropic-hosted deployments, data may be processed outside Azure and your selected region, under Anthropic's data processing terms — a point your security review will care about.

If your workload needs…On FoundryPractical route
Messages, streaming, tool useGAUse Foundry as-is
Batch processing at discountNot availableFirst-party API or Claude Platform on AWS
Files API, structured outputs, server-side toolsBeta, Anthropic-hosted onlyDeploy the Anthropic-hosted model version; review data-flow implications
Rate-limit headers, Admin/Usage APIsNot availableAzure Monitor, Cost Management, RBAC instead

"When will X arrive?" — how to actually track it

Neither Microsoft nor Anthropic publishes a feature roadmap for Claude on Foundry, so treat any expected dates you hear as rumor. What you can do is watch the authoritative status surfaces: Anthropic's Claude-on-Foundry page (which maintains the supported/unsupported lists quoted here), Anthropic's platform availability matrix, and Microsoft Learn's Claude model pages, which record per-model, per-hosting-option capabilities and lifecycle stages. The general pattern visible in the docs is that new capabilities tend to reach the Anthropic-hosted option first and Azure-hosted later — so if a missing feature is blocking you, the Anthropic-hosted model version is the first thing to test.

Rule of thumb: design the workload against the GA core, isolate anything beta or Anthropic-hosted-only behind an interface, and re-check the availability tables at every planning cycle. The gap list this article describes is a snapshot, not a constant.

Where to go next

See beta feature caveats for what the beta label costs you in practice, and the feature matrix for how Foundry compares with Bedrock, Vertex AI, and Claude Platform on AWS.

Sources