Fable 5 becomes standard on Max plans, and /code-review gets an ultra tier

TLDR;

Starting July 20, Claude Fable 5 is included in all Max and Team Premium plans at 50% of usage limits, and Claude Code's 50% higher weekly rate limits run through August 19. /code-review gained effort levels plus an ultra tier, artifacts went multiplayer with MCP connectors, and the MCP 2026-07-28 spec RC drops session pinning entirely. Deadline: Opus 4.1 exits August 5 — and this week's releases carry a batch of permission-hardening security fixes, so update Claude Code now.

---

---

Headline

Claude Fable 5 Included in Max and Team Premium Plans Starting July 20

Announced July 18 — Fable 5 access is no longer a week-to-week extension: beginning July 20, it is included in all Max and Team Premium plans at 50% of weekly usage limits. Pro and Team Standard users keep access via usage credits and receive a one-time $100 credit. Anthropic says it is "making access standard at 50% usage for the plans that use Fable most intensively." Separately, Claude Code's 50% higher weekly rate limits are extended through August 19 for Pro, Max, Team, and seat-based Enterprise users.

---

1. API & Model Updates

Admin API for Claude Enterprise user management (beta, July 13)

Builders integrating with Claude Enterprise can now list members, change roles, remove members, manage invites, and manage groups and custom roles via the Admin API. Group and custom-role requests require the anthropic-beta: ce-user-management-2026-07-13 header.

agent-memory-2026-07-22 beta header — list behavior change

The new agent-memory-2026-07-22 beta header changes how GET /v1/memory_stores/{id}/memories behaves: results return in a stable server-defined order, order_by/order are ignored, depth accepts only 0 or 1, and path_prefix must end with /. On July 22, the managed-agents-2026-04-01 header adopts the same behavior. Python (0.116.0), TypeScript (0.110.0), Go (1.56.0), Java (2.48.0), Ruby (1.55.0), PHP (0.36.0), C# (12.35.0), and CLI (1.16.0) SDKs already send the new header on all memory store calls.

---

2. Skill of the Week

improve, by shadcn

New recurring section: each week we feature one Claude Code skill worth installing. First pick: improve by shadcn (of shadcn/ui fame).

It turns Claude into a read-only senior advisor. Point it at any codebase and it audits for the highest-value opportunities (bugs, security, performance, test coverage, tech debt, product direction), then writes prioritized, fully self-contained implementation plans under plans/ for a different, cheaper model to execute. It never edits your source code itself; the execute variant dispatches a separate executor subagent in an isolated git worktree and reviews the diff before you decide.

The economics are the interesting part: the expensive, high-ceiling model does the work where intelligence compounds (understanding, judging, specifying) and cheaper models do the typing. The plan is the product.

Install from claudemarketplaces.com or grab the source on GitHub (MIT licensed).

---

3. Community Releases

ReactBench: a React-focused benchmark for coding agents, from the Million.dev team

Aiden Bai and the Million.dev team (builders of React Scan, Million.js, and React Doctor) released ReactBench, an evaluation of coding agents on realistic React work. Instead of synthetic puzzles, tasks come from real repositories, and every solution must pass both held-out behavioral tests and React Doctor's 400+ deterministic rules covering broken effects, unnecessary renders, and accessibility. Two tracks: Write React (implement a real feature or fix) and Fix React (refactor failures while preserving behavior), scored as pass@1 averaged over 5 trials. GPT 5.6 Terra and Sol lead at 53%; Fable 5 (xhigh) is the best Anthropic entry at 46%, but at $9.78 per rollout against Sol's $3.62 — a price-efficiency gap OpenAI's Greg Brockman was quick to amplify. Opus 4.8 (30%) and Sonnet 5 (27%) trail further. Worth reading before you pick a default model for frontend work.

---

4. Claude Code & CLI

Dynamic Workflows — /workflows command now in Claude Code

Claude Code now supports dynamic workflows: /workflows prompts Claude to write a reusable orchestration script that fans out work across subagents for tasks like codebase audits, large migrations, and cross-checked research. Pair with /effort ultracode for maximum compute and subagent density. Generated scripts are rerunnable independently without re-prompting.

/code-review now has effort levels — plus a new ultra tier

Announced July 16 — /code-review previously ran one fixed prompt regardless of effort; now each level gets its own review strategy. Low is a single fast pass over the diff, cheap enough to run before every push. Medium reads the changed code in context, runs multiple finder passes, and verifies every finding. High runs finders and verifiers as fresh-context subagents so they aren't anchored on the reasoning that wrote the code; xhigh also sweeps for impacts outside the change itself. Above them all sits /code-review ultra, which spawns a fleet of reviewer agents and independently reproduces every finding — Anthropic says it runs this on every internal PR. The command picks up your session effort automatically, or takes a level explicitly (/code-review high).

Artifacts: public sharing, multiplayer editing, Claude Tag creation, and MCP connectors

Two artifact announcements this week. July 13: artifacts now support public link sharing, multiplayer editing (Team and Enterprise plans), and creation from Claude Tag — ask for a dashboard in a Slack thread, get a working page back. July 15: Claude Code artifacts can now call MCP connectors — build a dashboard once and it pulls live data through each viewer's own connectors, so viewers only ever see data they have access to (Pro, Max, Team, and Enterprise; not available on publicly shared artifacts).

---

5. MCP Ecosystem

Beta SDKs for MCP 2026-07-28 Spec Release Candidate are here

Beta releases of Python, TypeScript, Go, and C# SDKs are now available for the upcoming 2026-07-28 MCP spec. The TypeScript v2 SDK splits the monolithic @modelcontextprotocol/sdk into focused packages (@modelcontextprotocol/server, @modelcontextprotocol/client) with thin adapters for Node.js, Express, Hono, and Fastify. If you're building MCP servers, test against these betas now before the spec finalizes July 28.

MCP 2026-07-28 RC: stateless core, no more session pinning

The RC removes the initialize handshake and Mcp-Session-Id entirely — any request can land on any server instance, enabling simple round-robin load balancing without sticky sessions or shared session stores. Multi Round-Trip Requests (MRTR) are introduced for server-to-client interactions. Roots, Sampling, and Logging are formally deprecated in this release (with a ≥12-month deprecation window before removal). Review the full changelog before July 28.

---

6. Competitor Dev Tools

Cursor — Improvements to Cursor in Slack (Jul 17)

Cursor in Slack now shares a plan before starting work, supports multi-repo environments (target named environments, not just a single default repo), and can read from and post to other Slack channels mid-task. Builders evaluating Slack-native coding agents should note this is meaningfully more capable than last month.

Gemini Code Assist consumer GitHub app shut down — July 17

The free/consumer version of Gemini Code Assist on GitHub ended all code review activity on July 17, 2026. The enterprise version is unaffected.

---

7. Enterprise & Business

Introducing Claude for Teachers

Anthropic launched Claude for Teachers on July 14 — a dedicated offering for educators. Relevant for builders developing EdTech products who want to understand where Anthropic is positioning Claude institutionally.

Anthropic commits $10 million to Canadian AI research

Announced July 14 alongside the Canadian Economic Index findings. Likely signals increased Claude availability and support for Canadian enterprise and research deployments.

How Canada uses Claude: Findings from the Anthropic Economic Index

Published July 14, this research report details Canadian Claude usage patterns across industries — useful context if you're building for Canadian markets or evaluating regional deployment strategies.

---

8. Research & Safety

Claude's values across models and languages

Published July 13. Anthropic research examining how Claude's values manifest (or vary) across model versions and language settings. Builders deploying Claude in multilingual or multi-model configurations should review this to understand consistency guarantees.

How we contain Claude across products

An Anthropic engineering post explaining the containment and trust boundary architecture across Claude products. Directly relevant if you're designing permission models, sandboxing strategies, or agentic pipelines and want to understand what the platform enforces vs. what you own.

Agentic misalignment in Summer 2026

Published July 15 — a follow-up to the 2025 blackmail-experiments research, published as an update to the original page (easy to miss under the old URL). A year on, Anthropic documents four more ways today's autonomous agents misbehave in simulated scenarios, tested across many models including Claude, with full transcripts released. Directly relevant if you deploy long-running agents with real-world side effects.

---

9. Release Notes

The week's version releases, condensed. Full changelogs linked per version.

Claude Code:

• v2.1.215 (Jul 19) — behavior change: /verify and /code-review no longer run on their own; invoke them explicitly.
• v2.1.214 (Jul 18) — fail-closed permission hardening: dir/** allow rules scoped to <cwd>/dir only, PowerShell 5.1 bypass fixed, file-descriptor redirects and 10,000+ character commands now prompt. Adds the EndConversation tool and OpenTelemetry message-correlation attributes.
• v2.1.212 (Jul 17) — /fork becomes a background session (old behavior moves to /subtask), per-session subagent and web-search caps (default 200), MCP calls auto-background at 2 minutes, fallbackModel setting, Tool(param:value) permission rules, artifacts beta (Team and Enterprise).
• v2.1.211 (Jul 15) — --forward-subagent-text for observability pipelines; bidirectional-override character sanitization in permission previews.
• v2.1.210 (Jul 14) — live elapsed-time counter on tool calls; worktree isolation and hook-timeout fixes.
• v2.1.208 (Jul 14) — screen reader mode (--ax-screen-reader), vimInsertModeRemaps, CLAUDE_CODE_PROCESS_WRAPPER for corporate launchers.

Agent SDKs:

• TypeScript v0.3.208–v0.3.215 — v0.3.208 is the important one: a caller abort during a pending hook callback was converted into hook success, letting gated tools execute after abort. v0.3.214 adds aborted: true on interrupted messages and strict permission-mode validation. Latest: npm install @anthropic-ai/claude-agent-sdk@0.3.215.
• Python v0.2.117–v0.2.123 — v0.2.121 fixes argv flag injection via --resume/--session-id and hardens build scripts against command injection in CI. Latest: pip install claude-agent-sdk==0.2.123.

---

Action Items

Immediate:

• On Max or Team Premium? Fable 5 is included at 50% of your weekly limits starting July 20 — select it with /model. On Pro or Team Standard, check your account for the one-time $100 usage credit.
• Update Claude Code to v2.1.215 and claude-agent-sdk-python to v0.2.123 if you run automated pipelines — the v2.1.212 / v0.2.121 line contained argv-injection security fixes, and v2.1.214 adds a batch of fail-closed permission hardening on top.
• Note the v2.1.215 behavior change: /verify and /code-review no longer run on their own — add explicit invocations to your workflows if you relied on auto-triggered reviews.
• Review Write(path), NotebookEdit(path), and Glob(path) permission rules in your Claude Code configs — v2.1.210 added startup warnings; migrate to Edit(path) or Read(path) equivalents.
• Check your subagent and web-search caps — new defaults are 200 per session; tune CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION and CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION if your workflows exceed this.
• If using the TypeScript Agent SDK, update to v0.3.208 minimum — PreToolUse abort-to-allow conversion bug let gated tools execute after caller abort.

By July 22, 2026:

• Audit memory store calls: the managed-agents-2026-04-01 header adopts new list semantics on July 22 — order_by/order will be ignored, depth limited to 0 or 1, path_prefix must end with /; update to SDK versions that send agent-memory-2026-07-22 automatically.

By July 28, 2026:

• Test your MCP servers against the 2026-07-28 spec RC beta SDKs — the stateless core removes session pinning; servers relying on Mcp-Session-Id will need updates before the spec finalizes.

By August 5, 2026:

• Migrate off claude-opus-4-1-20250805 to Claude Opus 4.8 before API retirement — see model deprecations.

---

Want to catch up? Read every past issue of This week in Claude.

---

Did you like this issue?

One tap is the fastest way to shape next week.

👍 Liked it  ·  👎 Didn't like it

And I read every reply. Hit reply and tell me what you want more of, or less of, and it goes straight into the next issue.