Microsoft Foundry in Practice

Beta Feature Caveats on Microsoft Foundry

Claude on Foundry has broad feature coverage — and as of July 2026, the models and most core capabilities are GA, but some features still carry a beta label. Knowing which capabilities are GA, which are preview, and what that label implies keeps production plans honest.

Claude 3P 101 · Updated July 2026 · Unofficial guide

"Preview" and "beta" are not decorative words in the Azure ecosystem. They signal that a capability may change, may carry different support expectations than generally available (GA) services, and may not be appropriate for workloads that demand stability guarantees. Claude in Microsoft Foundry is unusual among the third-party Claude platforms in exactly this respect: the feature list is long, and while most of it is now GA, several items still carry beta labels.

What's actually GA versus beta

Per the availability data Anthropic publishes for its platforms, the capabilities that are GA on Foundry include the Messages API with streaming and tool use, fine-grained tool streaming, PDF input, prompt caching, adaptive and extended thinking, citations (including search_result blocks), and the 1M-token context window. A number of features still carry a beta mark on Foundry even where they are GA on other platforms — including structured outputs, token counting, web search and web fetch, code execution, tool search, the bash/text-editor/memory client tools, Agent Skills, programmatic tool calling, the MCP connector, and the Files API.

Beta status shows up at three distinct levels, and it pays to keep them separate:

What beta means for your risk register

Betas can change or retire on their own schedule. Foundry already has a concrete precedent: the Sonnet 4.5 1M-context beta (header context-1m-2025-08-07) was retired on April 30, 2026 — since May 1, requests over 200K tokens with that header on Sonnet 4.5 return an error. Any workload built on a beta needs an owner watching for exactly this kind of notice, and a fallback plan (in that case: move to a model where 1M context is standard).

Betas complicate support conversations. When you escalate an issue on a beta capability, expectations differ from GA. Capture the request-id and apim-request-id response headers on every call — Anthropic asks for both to trace issues across Anthropic and Azure systems — so at least the diagnostics are production-grade even when the feature is not.

Betas deserve feature flags. Wrap beta-dependent code paths behind your own flags so a behavior change upstream becomes a config toggle rather than an incident. This is standard advice; on Foundry it applies to an unusually large share of the surface area.

Rule of thumb: build production-critical paths on the GA core (Messages, streaming, tool use), treat every beta capability as an enhancement you could disable tomorrow, and revisit the labels quarterly — beta-to-GA promotion is steady, and a caveat that was true at design time may be gone by launch.

How to track promotions to GA

Three watch points cover it. First, Anthropic's Claude-on-Foundry documentation and platform availability tables, which mark each feature GA/beta/unavailable per platform. Second, Microsoft Learn's Claude model pages, which carry per-model, per-hosting-option lifecycle status and beta annotations. Third, model lifecycle notices: Foundry follows the Claude API model deprecation schedule (for example, Claude Opus 4.1 on Foundry retires August 5, 2026), so Anthropic's deprecation announcements apply to your Azure deployments too. Assign someone to check all three on a cadence; none of them will push a notification to you by default.

Where to go next

The companion piece, Foundry vs. Direct API: the current feature gap, lists what is missing outright rather than merely beta. For platform-wide context, see the feature matrix.

Sources

← All articles