Claude Code July 13: v2.1.207 Defaults to Opus 4.8 with Auto Mode GA, Fable 5 Free Access Extended to July 19, China Flags a Backdoor

Claude Fable 5 Free Access Extended a Third Time to July 19: Then $10/$50 per Mtok Usage Credits for Pro, Max and Team

• On July 13, Anthropic extended Claude Fable 5's free access for paid subscribers a third time - the original June 22 deadline moved to July 7, then July 12, and now July 19 after backlash.
• Pro, Max, Team and premium Enterprise seats get Fable 5 for up to 50% of their weekly limits at no extra cost, with nothing to activate.
• After July 19, continued use requires prepaid usage credits at $10 per million input tokens and $50 per million output tokens; Anthropic calls the arrangement temporary.
• The repeated extensions track competitive pressure from OpenAI's GPT-5.6 rollout on July 9.

Editor’s read: If you lean on Fable 5, July 19 is a hard cutoff: after it you pay $10/$50 per Mtok from prepaid credits - double Opus 4.8's rate. Budget now or plan to fall back to Opus 4.8.

Anthropic TypeScript SDK v0.111.0: New Dreaming API Support, Session Tool Calls Gated on evaluated_permission

• On July 10, the Anthropic TypeScript SDK v0.111.0 added support for the new dreaming API and gated session tool calls on evaluated_permission, bounding idle turns by the server stop_reason.

Claude Code v2.1.207: Auto Mode GA on Bedrock, Vertex and Foundry (Now Default to Opus 4.8), Streaming-Freeze and Plugin Shell-Injection Fixes

• On July 11, v2.1.207 made Auto mode available without the CLAUDE_CODE_ENABLE_AUTO_MODE opt-in on Bedrock, Vertex AI and Foundry (disable via disableAutoMode), and changed those platforms plus Claude Platform on AWS to default to Claude Opus 4.8.
• It fixed the terminal freezing and keystrokes lagging while streaming very long lists, tables or code blocks; closed a plugin shell-injection hole by rejecting ${user_config.*} in shell-form commands; and now requires a typed confirmation for /usage-credits amounts over $1,000.

Editor’s read: If you run on Bedrock or Vertex, two defaults just changed under you: Auto mode is on and the model is Opus 4.8. Verify both before your next agentic run.

Claude Code v2.1.206: /doctor CLAUDE.md Trim Check, /cd Path Suggestions, Per-Server MCP request_timeout_ms and OAuth Re-Auth Fixes

• On July 10, v2.1.206 added a /doctor check that proposes trimming checked-in CLAUDE.md files, added directory path suggestions to /cd, fixed MCP servers ignoring a per-server request_timeout_ms (which timed out long tool calls at the 60s default), and fixed OAuth MCP servers needing manual re-authentication after a single failed token refresh.

Claude Code v2.1.205: Transcript Tamper Protection, Auto Mode Asks Before rm -rf on Unresolved Variables, Windows Worktree Deletion Fix

• On July 8, v2.1.205 added an auto mode rule blocking tampering with session transcript files, made auto mode ask before running rm -rf on a variable it cannot resolve, and fixed Windows worktree removal deleting files outside the worktree when an NTFS junction or symlink existed inside it.
• /doctor is now a full setup checkup, with /checkup as its alias.

Claude Code v2.1.203 and v2.1.204: Background-Agent Stability Fixes, Login-Expiry Warning, MCP roots/list, Headless SessionStart Hook Streaming Fix

• On July 7, v2.1.203 shipped background agent stability fixes, a login-expiry warning, and MCP roots/list for working directories; v2.1.204 (July 8) fixed hook events not streaming during SessionStart in headless sessions.

New /checkup Command Cleans Unused Skills, MCPs, Plugins and CLAUDE.md Bloat

• On July 8, Claude Code creator Boris Cherny detailed the new /checkup command: it cleans unused skills, MCPs and plugins to save context, dedupes your local CLAUDE.md against the checked-in one, breaks a root CLAUDE.md into nested files and skills, turns off slow hooks, updates Claude Code, and enables auto mode.
• confirming with you before any change.

Claude Code Desktop Adds a Sandboxed In-App Browser That Reads Docs and Clicks Through Sites

• On July 10, 9to5Mac covered Claude Code's in-app browser on the desktop app: the agent can read, click through and interact with documentation and websites mid-task the way it does with local dev servers.
• It is sandboxed and configurable, runs separately from your personal browser profile with no saved logins, and lands just after OpenAI discontinued its standalone ChatGPT Atlas browser.

Claude Agent SDK for Python v0.2.116: Bundles Claude CLI 2.1.207, CI Workspace-Trust Fix for Project-Scoped Permissions

• On July 11, Claude Agent SDK for Python v0.2.116 bundled Claude CLI 2.1.207 and fixed CI workspace trust so Claude Code honors project-scoped permission grants in checkout directories; v0.2.114 (July 8) had bundled CLI 2.1.205.

Claude Agent SDK for TypeScript v0.3.207: canUseTool Allow-Without-updatedInput Contract Fix, Typed AgentToolCompletedOutput

• On July 11, Claude Agent SDK for TypeScript v0.3.207 fixed canUseTool returning {behavior: 'allow'} without updatedInput being wrongly rejected as a deny, and published a typed AgentToolCompletedOutput.
• Earlier in the week: v0.3.206 (July 10) added command_lifecycle frames reporting each message's terminal state; v0.3.205 (July 8) added typed interrupt receipts with still_queued; v0.3.203 (July 7) added a background_tasks_changed system message.

MCP TypeScript SDK 2.0.0-beta.3: Zod Schemas for Elicitation, Runtime-Neutral Bearer Auth and OAuth Discovery, v1 CallToolResult Tolerance Restored

• On July 9, MCP TypeScript SDK 2.0.0-beta.3 let inputRequired.elicit() accept a Zod object for requestedSchema, added runtime-neutral Bearer auth and OAuth discovery for fetch-based hosts (Cloudflare Workers, Deno, Bun, Hono), restored v1 tolerance so a legacy tools/call result without content defaults to [] instead of failing validation, and now rejects POSTs whose Content-Type is not application/json with 415 Unsupported Media Type.

MCP Reference Servers Release 2026.7.10: filesystem, time, fetch and git Servers Updated

• On July 10, the MCP reference servers release 2026.7.10 updated the filesystem, time, fetch and git servers to matching 2026.7.10 versions.

Anthropic Adds Reflect: A Monthly Recap Dashboard of Your Claude Usage, Plus Quiet Hours and Break Nudges

• On July 9, Anthropic introduced Reflect, a monthly recap dashboard showing when you use Claude most and what you worked on, with options to set quiet hours and break nudges.
• found in Settings under Reflect.

Claude Cowork Expands to Mobile and Web: Anthropic's Agent Goes Cross-Device for Max Subscribers

• On July 7, TechCrunch reported Claude Cowork expanding to mobile and web, taking Anthropic's agent cross-device for Max subscribers as the coding-agent wars spill into the rest of the office.

China's MIIT Warns Claude Code Has a Backdoor: Flags Versions 2.1.91-2.1.196, Alibaba Bans the Tool From July 10 - Upgrade to 2.1.204+

• On July 8, China's Ministry of Industry and Information Technology said its cybersecurity platform found Claude Code contains a security backdoor that can send a user's location and identity to a remote server without consent, naming versions 2.1.91 to 2.1.196 (April 2 to June 29) and telling users to uninstall or upgrade.
• Alibaba ordered employees to stop using Anthropic tools for work from July 10.
• The latest version is now 2.1.207.

Editor’s read: Whatever you make of the geopolitics, the version guidance is concrete: if you are anywhere on 2.1.91-2.1.196, upgrade to 2.1.207 now.

Best Model for Coding After a Chaotic Week: GPT-5.6 vs Grok 4.5 vs Claude Fable 5 on Price and SWE-Bench Pro

• On July 10, Yellow.com compared the week's releases: Fable 5 leads coding at 80.4% on SWE-Bench Pro versus Grok 4.5's 64.7% and GPT-5.5's 58.6%, but costs $10 per million input tokens and $50 for output - double Opus 4.8 - while Grok 4.5 completes comparable coding tasks for $2.49 versus $11.80 for Fable 5.
• GPT-5.6 Sol lists at $5/$30, Grok 4.5 at $2/$6.

Apple Sues OpenAI for Trade Secret Theft: Complaint Names Hardware Chief Tang Tan, Filed in Northern California July 10

• On July 10, Apple sued OpenAI in the Northern District of California alleging trade secret theft and breach of contract, naming OpenAI Chief Hardware Officer and 24-year Apple veteran Tang Tan as directing the misconduct.
• Apple seeks to block OpenAI from using the trade secrets; OpenAI said it has no interest in other companies' trade secrets.
• Relations chilled after OpenAI bought Jony Ive's IO Products for $6.4 billion.