Five betas go GA on August 19, Claude Academy launches with 26 courses, Fable 5 biology safeguards retuned

Computer use reached general availability as computer_toolset_20260801 — no beta header required on the Claude API, also available in beta on Amazon Bedrock and Google Cloud Vertex AI. Alongside it, browser use shipped as an entirely new toolset (browser_toolset_20260801), not a renamed version of computer use. The distinction matters: computer use reads pixels and controls a desktop via screenshot; browser use reads a page's DOM structure and drives a browser your application hosts. Both are HIPAA-compliant. Computer use now supports multi-action execution in a single model call. Skills API and Files API are available on Microsoft Foundry; browser use and computer use on Vertex AI are coming soon.

The Files API (/v1/files) is now generally available. Drop the files-api-2025-04-14 beta header — requests without it use the GA response format, which adds file expiration (set expires_in_seconds on upload; objects report expires_at), page and next_page pagination, and an ids[] filter on list calls. Files API now offers 5x higher rate limits and 1 TB of storage per organization. Agent Skills and the Skills API (/v1/skills) are also GA — drop the skills-2025-10-02 beta header. The Enterprise Admin API user-management endpoints (members, invites, groups, custom roles) are GA as well — the anthropic-beta: ce-user-management-2026-07-13 header is no longer required. Managed Agents gained allowed_domains and blocked_domains controls for web tools, and self-hosted-sandbox sessions can now attach memory stores.

Anthropic launched Claude Academy, a free learning hub at academy.claude.com, on August 20 — the day after the big GA release, underscoring that the platform and education stories are intentionally moving together. The platform opens with 26 self-paced courses organized into four tracks: getting started, AI fluency for non-technical professionals, building with the Claude API, and Claude Code for developers. Courses award completion certificates and badges. A dedicated Claude Academy Skill is available for Claude Code. Anthropic says millions of people already visit anthropic.com monthly to learn AI and describes the Academy as its scaled answer to that demand. No credit card required.

Anthropic published the Claude Code Guide for Startups, distilling five operating rules drawn from more than a dozen real Claude Code deployments at early-stage companies. The guide covers how to scope tasks, set permission defaults, structure CLAUDE.md files, manage token budgets, and build review gates — practical operational advice rather than a feature walkthrough. Available on the Claude blog.

Anthropic retuned Fable 5's biology safeguards following significant feedback that the original classifier was over-blocking legitimate research queries. The retuned classifier reduces blocked biology queries by 85% and cuts total Claude.ai Fable 5 fallbacks to Opus 4.8 by 67%. Anthropic says the improvement was achieved without reducing the classifier's effectiveness against genuinely dangerous biosecurity requests, citing internal red-team results. This follows the June 11 controversy over the invisible AI research safeguard, and is part of Anthropic's broader effort to calibrate visible safeguards more precisely.

Starting August 14, new Claude Code sessions on Pro, Max, and Team plans default to auto mode rather than manual. Anthropic says its auto mode safety classifier now catches 89% of dangerous commands compared to 13.6% before the June safety upgrades. Existing users who had previously set a preference retain it; the default only applies to fresh installs and new sessions with no stored preference. Teams that prefer manual can set it via /config or the disableAutoMode managed setting.

The experimental prompt tools APIs — /v1/experimental/generate_prompt, /v1/experimental/improve_prompt, and /v1/experimental/templatize_prompt — were retired on August 17 alongside the legacy Workbench. Requests to these endpoints now return an error. Teams using these endpoints for prompt generation or improvement should migrate to the Prompt Improver in the Claude Console or equivalent manual prompting workflows.

Claude suffered two separate disruptions this week. On August 16 at 21:58 UTC, an authentication failure prevented users from signing in to Claude.ai, Claude Code, and Claude Cowork; broader degraded performance across Claude.ai and platform.claude.com followed within minutes. A second, smaller degraded performance incident hit on August 18 around 12:25 PM Eastern Time. Both incidents resolved within a few hours. These follow the August 5 outage (7.5 hours, all flagship models) and continue a reliability trend that has been a recurring topic in developer communities ahead of Anthropic's anticipated IPO.

Claude Sonnet 5's introductory pricing of $2 per million input tokens and $10 per million output tokens expires on August 31, 2026. Standard pricing of $3/$15 per MTok — unchanged from Sonnet 4.6 — takes effect September 1. Note: Sonnet 5's updated tokenizer maps the same content to 1.0–1.35x more tokens than previous versions, so the effective cost increase on identical workloads will be larger than the list-price change alone suggests. Audit your Sonnet 5 token counts before September to avoid budget surprises.