Five Betas Went GA. Which Headers Can You Delete Now

The Claude Developer Platform had a maintenance day on August 19, 2026 — and for teams running Claude integrations in production, that is better news than a feature day. Five capabilities reached general availability at once: computer use, browser use, the Files API, Agent Skills with the Skills API, and Admin API user management for Claude Enterprise.

The instinctive question for any integration owner reading a GA announcement is “what breaks, and by when?” Here the answer is unusually calm: nothing breaks, and no deadline exists. Every beta header these GA moves supersede is still accepted for backward compatibility, earlier beta tool versions remain available, and — after checking both Anthropic’s general beta-header policy page and each individual release note — we can say no sunset date is published for any of them.

This is not another how-to. We already publish implementation guides for computer use and Agent Skills, and we cross-link them below. This post does the maintenance-day job instead: exactly what changed, which headers you can delete, what silently keeps working if you do nothing, and the one place where upgrading changes behavior underneath a same-looking call.

The August 19 release-notes entry reads as five bullets, but they are not five equal stories. Four of them retire beta headers or beta tool versions for capabilities teams already run in production. One — browser use, shipping as browser_toolset_20260801 — is a genuinely new tool that drives a browser your application hosts through its accessibility tree, and it deserves its own treatment: our deep dive on the new browser use tool covers it in full, so here it gets exactly one row in the table below.

The table is the checklist version of the whole event: what the old identifier was, what the GA identifier is, what actually changes beyond the header, and whether you need to do anything this quarter.

One framing note before the detail: the pattern across all five moves is identical. Old headers are accepted, not rejected. A request that still sends files-api-2025-04-14 keeps working and returns the previous response format; a request that still sends the Skills or user-management header behaves identically to one that does not. That single design decision is what turns a five-item changelog into a calm afternoon instead of a sprint-planning emergency.

Computer use leaves beta on the Claude API as the computer_toolset_20260801 toolset. The headline is the missing beta header, but the more consequential story is that this is a structural revision, not a rubber stamp. Where the beta tool was one monolithic tool with an action field, the GA toolset exposes 17 member tools for screenshot, mouse, and keyboard control of a desktop environment, per the Computer Use Tool reference.

Three of those clauses deserve unpacking, because each one changes what your integration code sees:

Batch actions: multi-action turns

Claude can now return several tool_use blocks in a single response — click, type, screenshot — without waiting for each intermediate result. Execution is strictly sequential, not parallel. The subtle new obligation is in error handling: if one block fails, execution stops there, but every subsequent block in that turn must still receive a tool_result, with content stating it was not executed because an earlier action failed. Harness code written for the one-action-per-turn beta has never needed to fabricate results for skipped actions; the GA toolset requires it.

Zoom: on by default

The zoom tool lets Claude request a cropped, full-resolution view of a screen region — useful for small text, dense UI, file names, or status-bar content — while the coordinates it uses afterward stay in full-screenshot space. Anthropic’s docs are explicit that zoom is enabled by default for all 17 member tools. It can be switched off per integration via configs if your environment cannot produce zoomed images, but the toolset ships with it on.

A new request/response shape

Upgrading is a real (if small) migration, documented in the reference’s own migration table: the display parameters (display_width_px, display_height_px, display_number) are no longer needed because coordinates are always in screenshot pixel space; tool_use and tool_result blocks now carry a toolset_name field and dispatch on the toolset-name plus tool-name pair instead of a single action field; and zoom moves from an enable_zoom parameter to a built-in member tool.

Two scoping caveats keep this honest. First, the no-beta-header behavior is Claude API only: as of the date checked, Claude Platform deployments on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry still require a beta header or only offer the earlier tool versions. Second, the GA toolset is supported on Claude Fable 5, Claude Mythos 5, Claude Opus 5, Claude Sonnet 5, and Claude Opus 4.8 — the same five-model list the release notes give for browser use.

None of this changes how you should build a computer-use agent — for that, our computer use API build guide and production deployment guide for computer use remain the implementation references, and our cross-vendor computer-use agent matrix covers the competitive picture. What GA changes is the operational posture: this surface is no longer experimental by Anthropic’s own labeling, and the toolset revision is the shape new work should target.

The Files API is GA on the Claude API: requests to the /v1/files endpoints, and Messages API requests that reference an uploaded file, no longer require the files-api-2025-04-14 beta header. Requests that still send it keep working and return the previous response format. But of the five GA moves, this is the one where dropping the header changes what comes back — which makes it the most concrete code-touch in the release, and the item most secondary coverage will likely skip past in favor of the flashier toolsets.

Pagination changes shape

Without the beta header, the list endpoint moves from cursor-only pagination to page / next_page paging, plus a new ids[] filter, per the Files API guide. The defaults: 20 files per page, 1,000 maximum, newest first. The ids[] filter is genuinely useful — check up to 100 known file IDs in one request without paging — but it carries a quiet gotcha: any ID that does not resolve is silently omitted from the response rather than raising an error, so callers must diff the returned IDs against the requested ones to detect misses. It always returns a single page (next_page comes back null) and cannot be combined with page or limit.

File expiration is new

GA also brings a real lifecycle feature: set expires_in_seconds at upload, from 3,600 seconds (one hour) to 7,776,000 seconds (90 days). The resulting expires_at timestamp is null if no expiration was set, and the value is fixed at upload — it cannot be changed afterward. Once expires_at passes, downloading the file’s content returns a 404 and a Messages request referencing it fails before inference runs. But metadata stays readable for up to 30 days past expiration, and the file keeps appearing in list responses during that window — so list-consuming code has to compare expires_at against the current time to filter expired files out itself.

The operational numbers worth pinning to your runbook: file storage caps at 500 MB per file and 1 TB per organization; file-related calls are rate-limited to roughly 500 requests per minute; and Files API operations themselves — upload, download, list, get metadata, delete — are free. Only file content used inside Messages requests is billed, as input tokens.

Agent Skills and the Skills API (/v1/skills) are GA on the Claude API. Requests — including Messages API requests that load Skills through the container parameter — no longer require the skills-2025-10-02 beta header, and requests that still send it continue to work unchanged. That is the entire delta: no behavior change, no new response shape. If you are building with Skills rather than just maintaining them, our Agent Skills framework guide is the authoring deep dive.

The Admin API move needs more precision, because it is easy to over-generalize. What reached GA is Claude Enterprise user management across all four resource types — members, invites, groups, and custom roles. The header change, though, is scoped: the ce-user-management-2026-07-13 beta header is no longer required specifically on group and custom-role requests. Member and invite endpoints never needed that particular header in the first place — they only ever required anthropic-version. So the accurate sentence is “user management is GA, and the one header it ever required is now optional,” not “the Admin API no longer needs any beta header.”

We checked this two ways, because it is the question every integration owner actually has. First, Anthropic’s general beta headers reference carries only generic policy language — beta features “may” have breaking changes with notice, be deprecated or removed, or carry different limits — with no dates and no per-feature sunset table. Second, each individual GA release note explicitly says the old header or version “remains available,” “continues to work unchanged,” or is “accepted unchanged,” again with no stated end date. As of the date checked, no sunset or deprecation date is published for computer_20251124, files-api-2025-04-14, skills-2025-10-02, or ce-user-management-2026-07-13.

Read that absence carefully in both directions. It does not mean these identifiers are supported forever — Anthropic’s own policy language reserves the right to deprecate with notice. It means the migration is yours to schedule: there is no externally imposed deadline forcing this into the current sprint. A useful mental model is the naming convention itself. Beta headers follow feature-name-YYYY-MM-DD, where the date marks when that beta surface was released — each one is a dated snapshot of an API surface, not a version number. GA froze those snapshots in place and made the current surface the default; it did not detonate them.

Our interpretation: this release pattern — additive identifiers, accepted legacy headers, no published sunsets — is Anthropic optimizing for enterprise trust at exactly the moment these capabilities move from experiments to production dependencies. Whether that posture holds across future releases is not something a single release-notes entry can establish, so the rational response is neither panic nor complacency: put the migration in next quarter’s backlog with a note to re-check the release notes for deprecation announcements before deprioritizing it again.

The August 19 entry carried three adjacent changes worth one clause each. Managed Agents sessions gain allowed_domains and blocked_domains configuration restricting which sites a session’s web_search and web_fetch tools can reach, set inside the agent_toolset_20260401 configs array — with web_fetch also gaining max_content_tokens and web_search gaining user_location. Managed Agents sessions running in a self-hosted sandbox can now attach memory stores, with the Python, TypeScript, and Go SDK workers downloading each attached store into the sandbox at its mount_path and syncing changes back. And the Claude Console’s session viewer was redesigned with a timeline minimap, per-model-request transcript grouping, and an Inspector panel.

For the Managed Agents backdrop — effort controls, webhooks, and how Skills plug into sessions — our Managed Agents update roundup covers the surface these domain controls now bolt onto.

With no deadline forcing the issue, the right move differs by what you run. Four common situations, four different answers:

Looking forward, the same-day GA of computer use and browser use is the signal worth projecting from: Anthropic now treats agent-environment control — desktop and browser — as a stable, supportable product surface rather than a research preview. Teams that deferred computer-use projects on “it’s still beta” grounds have lost that reason, and the sensible next step is a scoped pilot rather than a platform-wide commitment. If you are deciding which of these surfaces belongs in your automation roadmap — and which workloads justify an agent that drives a screen at all — our AI transformation engagements start with exactly that evaluation.