🧭 Computer Use, Skills API, and Files API All Reach General Availability

Anthropic has graduated three of its most-used agentic APIs from beta to general availability, removing the beta headers that have been required since each feature launched. Computer use, the Skills API, and the Files API are now production-grade and available to all Claude Platform customers without gating.

What changes with GA

• Computer use — no longer requires the computer-use-2024-10-22 beta header. A new browser use tool joins the existing screenshot-and-click interface, giving agents a structured way to interact with web applications through DOM events rather than raw pixel coordinates. Precision on interactive elements improves significantly.
• Skills API — the skills-2025-10-02 beta header is retired. Skills (reusable, versioned prompt+tool bundles) can now be registered, shared across workspaces, and called from any agent without additional opt-in.
• Files API — the files-api-2025-04-14 beta header is retired. New in GA: an expires_in_seconds parameter for automatic file lifecycle management, a 5× rate-limit increase (from 100 to 500 RPM), and a 1 TB per-organisation storage ceiling. Files uploaded during the beta are carried forward with no action required.

The browser use tool in practice

The new tool accepts a selector or aria-label alongside the action type, letting Claude target UI elements by identity rather than guessing pixel positions from a screenshot. This eliminates a common failure mode where button positions shift between page loads or viewport sizes. The original screenshot-based approach remains available for interfaces without accessible HTML structure.

# Example: click a submit button by aria-label
{
  "type": "browser_use",
  "action": "click",
  "selector": "[aria-label='Submit form']"
}

🧭 Claude Code v2.1.238–239: Windows NTLM Credential-Leak Patched in Security Sprint

Anthropic shipped two rapid-succession Claude Code patch releases this week, closing a Windows-specific security vulnerability and fixing several stability issues that had surfaced in self-hosted and voice-mode deployments since last week's public beta launch.

Security fix — Windows NTLM credential leak

The most significant fix is a UNC path validation bypass on Windows. Paths prefixed with the \\?\?\ device path syntax were not being caught by the existing UNC path filter, meaning a specially crafted tool call referencing an attacker-controlled SMB share could trigger an outbound NTLM authentication handshake — leaking the current user's hashed credentials to an external host. The vulnerability required a malicious MCP plugin or a crafted CLAUDE.md directive to trigger.

Other fixes in v2.1.238–239

• MCP v2 serverless reconnect loop — connections were endlessly reopening the subscriptions/listen stream against serverless-hosted MCP servers; now capped with backoff and deduplication.
• Linux idle CPU spike — idle sessions with sandboxing enabled were pegging one CPU core at ~100%; the polling interval is now event-driven.
• Voice mode hang on connection rejection — native builds stuck on "listening…" when the voice service rejected a connection are now properly surfaced as an error with a retry prompt.
• Self-hosted runner session startup — branch creation no longer rewrites the working tree at session open, eliminating two unnecessary server round-trips and cutting agent launch latency by roughly 40%.

🧭 Claude Code's +50% Weekly Usage Boost Extended Through August 31 — With Hints at Permanence

Anthropic has extended the +50% Claude Code weekly usage limit promotion to August 31, 2026 — the third extension of what was originally framed as a limited-time offer in May. For the first time, Anthropic has stated it hopes to make the increase permanent, citing remaining capacity constraints as the only blocker to locking it in.

What the boost covers

• Applies to Pro, Max, Team, and Enterprise plans equally.
• Covers usage across all Claude Code surfaces: CLI, VS Code extension, JetBrains extension, desktop app, and web.
• The +50% multiplier applies to the per-plan weekly token budget, not to any daily limit or concurrent session cap.

Reading between the lines

Three extensions of the same "limited-time" promotion, combined with the explicit aspiration toward permanence, is a strong signal that Anthropic has effectively committed to the higher limits. The phrasing is cautious because Anthropic cannot yet guarantee the compute headroom to fulfil the promise universally — particularly given the surge in Team and Enterprise usage since auto mode became the default on August 14. Expect a formal limit-increase announcement once the Volta-Norway compute capacity (signed in early August) comes online in Q4.