Claude Code July 20: Fable 5 Included in Max Plans at 50% Limits, v2.1.214 Hardens Bash Permission Checks, /verify No Longer Auto-Runs

Claude Fable 5 included in all Max and Team Premium plans at 50% limits from July 20, one-time $100 credit for Pro

• On July 18, Anthropic announced that beginning July 20 Claude Fable 5 is included in all Max and Team Premium plans at 50% of usage limits, ending weeks of repeatedly extended free-access deadlines.
• Pro and Team Standard users keep access through usage credits and receive a one-time $100 credit.

Editor’s read: The stage-by-stage rollout is over - if you run Fable 5 daily, Max is now the plan that includes it by default.

Claude Code v2.1.214 hardens Bash permission checks: fail-closed redirects, PowerShell 5.1 bypass fix, new EndConversation tool

• On July 18, v2.1.214 shipped a security-focused batch: Bash permission checks now fail closed on file-descriptor redirect forms, commands over 10,000 characters always prompt, a permission-check bypass in Windows PowerShell 5.1 sessions is fixed, and single-segment allow rules like Edit(src/**) no longer auto-approve writes to nested directories anywhere in the tree.
• Claude also gains an EndConversation tool for ending sessions with highly abusive users or jailbreak attempts.

Editor’s read: If you rely on allow rules for autonomy, re-test them on 2.1.214 - several previously auto-approved command shapes now prompt.

Claude Code v2.1.215: /verify and /code-review skills no longer auto-run, invoke them manually

• On July 19, the latest release stopped Claude from running the /verify and /code-review skills on its own; invoke them explicitly with /verify or /code-review when you want a verification or review pass.

Claude Code v2.1.212: /fork copies your session to a background agent, WebSearch and subagent spawns capped at 200 per session

• On July 17, v2.1.212 changed /fork to copy the conversation into a new background session while you keep working (the old in-session behavior is now /subtask), moved MCP tool calls running longer than 2 minutes to the background automatically, and added per-session caps of 200 WebSearch calls and 200 subagent spawns to stop runaway loops.
• It also fixed plan mode auto-running file-modifying Bash commands like touch and rm without a permission prompt.

Claude Code v2.1.211 forwards subagent text in stream-json and neutralizes look-alike characters in permission previews

• On July 15, v2.1.211 added a --forward-subagent-text flag to include subagent text and thinking in stream-json output, fixed permission previews so bidirectional-override, zero-width and look-alike quote characters cannot visually alter the approval message, and floored a PreToolUse hook's ask decision at a prompt even in auto mode.

Claude Code v2.1.210 fixes a worktree-isolation escape and hardens prompt-injection surfaces

• On July 14, v2.1.210 fixed isolation: 'worktree' subagents being able to run git-mutating commands against the main repo checkout, stopped the ultracode keyword opt-in from firing on non-human input such as webhook payloads and relayed PR comments, and added a live elapsed-time counter to long-running tool calls.

Claude Code v2.1.208 adds screen reader mode, vim insert-mode remaps and major stability fixes

• On July 14, v2.1.208 added opt-in plain-text rendering for screen reader users (claude --ax-screen-reader), a vimInsertModeRemaps setting for two-key sequences like jj to Escape, and CLAUDE_CODE_PROCESS_WRAPPER for corporate launchers, alongside memory-leak and background-session stability fixes.
• v2.1.209 followed the same day to unblock /model and other dialogs in claude agents background sessions.

Agent SDK security fix: argv flag injection via resume and session_id patched in TypeScript 0.3.212 and Python 0.2.121

• On July 17, both Agent SDKs fixed dash-prefixed resume and session_id values being passed as separate argv tokens: --resume and --session-id are now single equals-joined tokens so a crafted value can never be misinterpreted as an independent CLI flag.
• The Python release also hardened build scripts against command injection via CLAUDE_CLI_VERSION.

Editor’s read: If session ids or resume targets in your agent ever come from user input, ship this update now.

Agent SDK TypeScript 0.3.214: set_permission_mode rejects unknown modes, aborted flag on interrupted messages

• On July 18, TypeScript 0.3.214 made set_permission_mode reject unrecognized permission modes with an error instead of silently adopting them, added aborted: true to assistant messages truncated by interrupt(), and added subagent_type and subagent_retry fields to tool_progress so clients can show a subagent waiting out a rate-limit retry.
• 0.3.215 followed on July 19 with parity for Claude Code v2.1.215, matched by Python 0.2.122 and 0.2.123 CLI bumps.

Agent SDK TypeScript 0.3.208 closes a PreToolUse abort gap

• On July 14, 0.3.208 fixed a caller abort during a pending SDK hook callback being converted into hook success, which let PreToolUse-gated tools execute after the abort.
• It also stopped a UserPromptSubmit hook timeout from killing the entire query and fixed dash-leading extraArgs values being parsed as their own CLI flags.

Agent SDK TypeScript 0.3.211 adds assistant message timestamps and rate-limit prefix buckets

• On July 15, 0.3.211 added SDKAssistantMessage.timestamp (ISO-8601) to the live stream, exported rate-limit message prefix buckets as alpha APIs for classifying usage-limit errors without hand-mirrored lists, and made process-exit errors include the CLI's stderr so a failed child reports its actual cause.

Anthropic TypeScript SDK 0.112.0 adds MCP Tunnels API support

• On July 16, the Anthropic TypeScript SDK shipped v0.112.0 with API support for MCP Tunnels, the new way to expose a local MCP server to Anthropic's API.

MCP TypeScript SDK 2.0.0-beta.4 moves schemas into @modelcontextprotocol/core, packages now version together

• On July 13, MCP TypeScript SDK 2.0.0-beta.4 moved the spec schemas, OAuth schemas and protocol constants into @modelcontextprotocol/core as a regular runtime dependency, so apps importing multiple MCP packages evaluate a single shared schema graph with shared object identity; the four packages now version together.

Anatomy of a misfeature: how Claude Code v2.1.198 silently shipped a 60-second AskUserQuestion auto-advance and v2.1.200 reversed it

• On July 17, developer Olaf Alders published a detailed post-mortem of a behavior silently introduced in v2.1.198 on July 1 and reversed three days later in v2.1.200: AskUserQuestion prompts would auto-advance with Claude's best judgment after 60 seconds, converting explicit safety gates into timeouts.
• The feature appeared in neither release notes nor docs when it shipped, and no public commit documented its addition or removal.

Editor’s read: A sharp case study in why auto-updating agent tooling needs changelog discipline - worth reading before you let unattended agents prompt humans.

Claude Code vs Codex vs Cursor vs Grok Build: the July 2026 state of play for AI coding agents

• On July 19, claude-world.com published a market snapshot arguing there is no universal winner - success depends on matching the right model-plus-harness combination to your workload, failure tolerance and scale.
• Anthropic fields four model tiers (Fable 5, Opus 4.8, Sonnet 5, Haiku 4.5) with subagents, Agent Teams, dynamic workflows and worktrees as parallelism options.

Claude Code, Codex and Cursor usage limits all reset on July 16

• On July 16, three major AI coding platforms refreshed quotas the same day: Anthropic reset Claude Code's 5-hour rolling window and weekly buckets, OpenAI refreshed Codex allocations, and Cursor doubled included usage for Grok 4.5 and Composer 2.5 on all paid tiers.
• framed as a retention war ahead of the Fable 5 promotional-access expiry.

Kimi K3: Moonshot AI ships a 2.8 trillion parameter model, the largest open-source release yet, and tops a major coding benchmark

• On July 18, coverage of Moonshot AI's Kimi K3 highlighted the largest open-weight model to date: 2.8 trillion parameters, native visual understanding and a one-million-token context window, placed behind Claude Fable 5 and GPT-5.6 Sol overall but with a 76% win rate on the Frontend Code Arena benchmark.
• Open weights are promised by July 27, 2026.

Gemini 3.5 Pro delayed again: misses its third launch deadline as Google eyes a stopgap Gemini 3.6 Flash

• On July 15, reports confirmed Gemini 3.5 Pro missed its widely reported July 17 general-availability date - its third slip since June - with reliability problems blamed and stopgap releases like Gemini 3.6 Flash under consideration.
• For developers, the delay leaves Claude, GPT-5.6 and Grok as the current frontier options for agentic coding.