Files, Agent Skills, and the Admin API leave beta

August 19 was beta-header removal day on the Claude platform. The Files API, Agent Skills with its /v1/skills endpoint, and the Admin API for managing users in Enterprise organizations all moved to general availability. Code you wrote with the anthropic-beta header still works, but the response shape in general availability is not identical to the one in beta.

What changes in the Files API

Requests to /v1/files, and Messages API calls that reference an uploaded file, no longer require the files-api-2025-04-14 header. Leave it off and you get the new response shape, which brings several additions.

• File expiry is set with expires_in_seconds at upload time, and the file object returns an expires_at field.
• Pagination arrives through page and next_page, with an ids[] filter when listing files.
• Storage is one terabyte per organization.
• The limit is 500 requests per minute on /v1/files.
• Requests that still send the beta header work unchanged and return the old response shape, so there is no forced migration.

Agent Skills without a beta header

Skills and the accompanying Skills API are generally available, so the skills-2025-10-02 header is no longer needed, including in Messages API calls that load skills through the container parameter. Skills stop being an experiment and enter the stable platform contract, with the same guarantees as the rest of the API.

In practice that means the same skill you keep locally in .claude/skills can run reliably through the API, without adding a header conditionally depending on the environment.

A fence around web access for managed agents

Anyone who lets an agent search the internet on its own should read this change. In the configs array, inside the agent_toolset_20260401 toolset, you can now set allowed_domains or blocked_domains, separately for search and for content fetching.

• The content fetching tool additionally takes max_content_tokens, an upper bound on how much content a single page may bring into the context.
• The search tool takes user_location, for results tied to a particular area.
• Each entry in the configs array is identified by name, and optionally by type.
• Earlier calls that send only name, enabled, and permission_policy keep working.

Smaller changes that still matter

• Sessions in a self-hosted isolated environment can now attach a memory store: a worker from the SDK downloads it to a given path and writes changes back.
• The session viewer in the console was redesigned and brings a timeline minimap, a transcript grouped by individual model request, and an Inspector panel with session data, cost, raw events, and per-tool statistics.
• Admin API endpoints for members, invites, groups, and custom roles are also generally available, with the earlier header still accepted.