TLDR;

Claude Opus 5 launched Friday at $5/$25 per million tokens — near Fable 5 intelligence at half the price, now the default on Claude Max, the default Opus model in Claude Code, and generally available in GitHub Copilot. The final MCP 2026-07-28 spec publishes tomorrow and takes the protocol stateless, with GitHub's MCP server already supporting it. Security had a heavy week: a fake Claude Desktop download page hosted as a real claude.ai Artifact hit at least 29 organizations, a newly disclosed symlink flaw in Claude Code remains unpatched, and Anthropic shipped a Claude Security plugin beta for Claude Code.

---

---

Headline

Introducing Claude Opus 5

Anthropic launched Claude Opus 5 on July 24 — a 1M-token-context model that lands "within 0.5% of Fable 5's peak score, but at half the cost per task," priced at $5/$25 per million input/output tokens (same as Opus 4.8). It scored 3x the next-best model on ARC-AGI-3 and a state-of-the-art 43.3% on Frontier-Bench v0.1. Opus 5 is the new default model on Claude Max, the strongest model on Claude Pro, and the default Opus model in Claude Code as of v2.1.219. A fast mode research preview runs at roughly 2.5x speed for $10/$50 per million tokens — Claude API only for now, and also usable inside Claude Code as extra usage. Alongside the launch, Anthropic published the new rules of context engineering for Claude 5 generation models, revealing they removed roughly 80% of Claude Code's system prompt for the Claude 5 generation — worth reading before you tune CLAUDE.md files or skills for Opus 5.

---

1. API & Model Updates

Claude Opus 5 for builders: model IDs, pricing tiers, and migration

The model ID is claude-opus-5 on the Claude API and Google Cloud, anthropic.claude-opus-5 on Bedrock; knowledge cutoff is May 2026. Four effective input price points off the same weights: cache read $0.50/M, batch $2.50/M, standard $5/M, fast mode $10/M. The Batch API supports up to 300K output tokens via the output-300k-2026-03-24 beta header. Claude Code ships a built-in /claude-api migrate skill that updates model strings and tunes prompts for Opus 5. And as of July 24, Opus 5 is generally available in GitHub Copilot on Pro+, Business, and Enterprise plans (admins must enable it).

Claude Managed Agents: effort levels, seeded sessions, and webhooks

The July 22 platform release adds per-agent effort levels, session seeding with up to 50 events at create time, a 500-skill-per-session ceiling, webhooks for environments and memory stores, and sub-agent event streaming — a meaningful batch if you run agents on Anthropic's managed infrastructure.

Voice mode grows up: Opus-powered, tool-connected, public beta

Claude's voice mode now runs on Opus and Sonnet models, can reach connected tools like email and calendar mid-conversation, and supports more languages — public beta across mobile, desktop, and web as of July 23.

---

2. Skill of the Week

draw-your-font: turn a photo of your handwriting into an installable font

Run npx skills add danilo-znamerovszkij/draw-your-font, drop in a photo of your handwriting, and Claude vision-labels the letters, builds a real TTF/WOFF/WOFF2 font locally (pure npm — no FontForge or ImageMagick), previews it, and iterates on feedback like "make it bolder." Open-sourced July 22 under MIT and already past 360 stars in five days, with a no-AI CLI fallback and a browser demo at drawyourfont.com.

---

3. Community Releases

Burp MCP Server (Unrestricted)

An unofficial fork of PortSwigger's official Burp MCP server rebuilt for long agent-driven security engagements: newest-first proxy history, site-map access, active scan and crawl triggers, and a full Repeater send/read loop, with the official version's 5,000-character response truncation made removable. Released July 25 under GPL-3.0 — small so far, but the README's capability table is verified against upstream source rather than marketing.

Rapid-MLX v0.11.0: reasoning-gated tool grammar for local models

The Apple Silicon local inference engine (3.4K stars, a drop-in OpenAI-compatible backend for Claude Code, Cursor, and Aider) shipped v0.11.0 on July 24: local models are now only forced into tool-call grammar after their thinking budget is spent, which keeps reasoning quality intact while claiming 100% tool-calling reliability across 17 tool parsers.

file-itr: a Claude skill that files Indian income tax returns

Filed under "skills are getting remarkably specific": this one prepares and e-files ITR-1 through ITR-4 on India's official tax portal and compares old vs new regimes to find the lowest legal tax. India-only, but a good preview of where vertical skills are heading.

---

4. Claude Code & CLI

Cowork "Record a skill": show Claude a task once, get a reusable skill

Announced July 21 (X-only so far): in the Cowork desktop app, hit the + menu, record your screen while doing a task and narrating it, and Claude turns the recording into a reusable skill. Available on Pro, Max, and Team plans. This is the lowest-friction skill authoring path Anthropic has shipped — worth trying on any workflow you repeat weekly.

Claude Code on desktop gets an iOS simulator (public beta)

Since July 21, Claude Code on macOS desktop can open an iOS simulator in a side panel (requires Xcode) — build, run, and iterate on iOS apps without leaving the session.

Nested subagents: disabled, then back at depth 3 — check your workflows

v2.1.217 (July 21) disabled nested subagent spawning by default (depth cap dropped from 5 to 1) and added a concurrent-subagent cap (default 20, override with CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS). v2.1.219 (July 24) re-enabled nesting at a default depth of 3, controlled by CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH. Separately, the advisory workflowSizeGuideline setting (unrestricted / small / medium / large, default medium at roughly 15 agents) is now settable from any settings file. If your orchestration relied on deep nesting, pin the env var explicitly rather than riding the shifting default.

How Anthropic runs large-scale code migrations with Claude Code

A July 21 deep dive published as an X Article — how Anthropic's own teams structure multi-repo migrations with Claude Code. Practical patterns if you're planning any large mechanical change.

---

5. MCP Ecosystem

MCP goes stateless tomorrow — and GitHub's server is already there

The final MCP 2026-07-28 specification publishes tomorrow, July 28 — the protocol goes stateless at the protocol layer, removing Sessions and initialize entirely. GitHub's MCP server already supports the new spec as of July 23, and all Tier 1 SDKs preserve backwards compatibility, so existing integrations keep working. New multi round-trip requests let remote servers support features like elicitation. If you maintain a server, note the beta SDKs: TypeScript v2 retires the monolithic @modelcontextprotocol/sdk package in favor of @modelcontextprotocol/server and @modelcontextprotocol/client.

---

6. Security

Claude Security plugin for Claude Code enters beta

Announced July 22: a Claude Security plugin that runs pre-commit vulnerability scans and full codebase scans straight from the terminal. If you've been wiring third-party scanners into Claude Code hooks, this is the first-party answer.

Fake Claude Desktop download page hosted on claude.ai itself — SectopRAT campaign

Between July 21–22, attackers published a public Claude Artifact rendering a convincing Claude Desktop download page on the real claude.ai domain, then ran Bing sponsored ads pointing at it. Huntress (writeup July 22) confirmed at least 29 organizations compromised via SectopRAT, delivered through a DLL-sideloading bundle with C2 details hidden in Ethereum transactions (EtherHiding). The artifact logged ~7,100 views before Anthropic removed it. Builder action: if your product surfaces Claude Artifact URLs to end users, call out the "Content is user-generated and unverified" banner explicitly in your UX — users reliably miss it.

New Claude Code symlink flaw disclosed — no fix yet

Disclosed July 27 by researchers at Tego: a flaw in Claude Code's startup memory loader follows symbolic links out of the repository, exposing readable files outside the project without explicit user approval — verified on v2.1.215, described as the third symlink-class flaw after CVE-2025-59829 and CVE-2026-25724. No CVE and no fixed version yet. Until a patch ships, review cloned repositories for unexpected symlinks before opening them in Claude Code, and watch the GitHub advisory feed.

---

7. Competitor Dev Tools

Cursor Router: Auto mode now routes across models (Jul 22)

Cursor's Auto mode now routes each request through Cursor Router with three optimization modes — Intelligence, Balance, and Cost — plus admin controls for defaults and model allow/block lists. On by default for Teams; Enterprise admins enable it from the dashboard. Available across desktop, web, iOS, CLI, and the SDK.

OpenAI Codex CLI v0.145.0: multi-agent V2 stabilized, /import adds Cursor (Jul 21)

The opt-in multi-agent V2 experience is now stabilized, with configurable sub-agent models, reasoning levels, and concurrency. /import — which has migrated settings, MCP servers, plugins, sessions, commands, and project memories from Claude Code since v0.140.0 — now imports from Cursor too. The portability play continues in both directions.

Zed v1.12.0: adaptive thinking for Anthropic models (Jul 23)

Zed adds adaptive thinking configuration for custom Anthropic models and turns on ACP elicitations by default.

---

8. Enterprise & Business

Claude Team plan minimum drops from 5 seats to 2

Announced July 20 (X-only): Team plans now start at 2 seats instead of 5 — a meaningful barrier drop for small teams and agencies that wanted central billing and admin without paying for empty seats.

Economic Futures: a research agenda and an Index connector

On July 22 Anthropic published the research agenda for its $200M Economic Futures Research Fund and shipped an Economic Index connector that lets any Claude conversation query the Index dataset directly — useful if you build labor-market or economic analysis features.

Anthropic donates another $20M to Public First Action

Announced July 21, bringing total support to $40M. Noted because Public First Action's policy work can influence AI regulation timelines that affect developer platform terms.

AI for Science rare disease grants — apply by August 2

Applications opened July 20: up to $50,000 in Claude credits over six months, two tracks (basic science and early-stage biotech for rare genetic diseases). Deadline August 2, 11:59 PM PST.

---

9. Research & Safety

Project Pilot: Can AI control a drone?

Published July 24: Anthropic's Frontier Red Team, with Andon Labs, evaluated 15 models from three developers on a locate-and-follow drone surveillance task. Claude Fable 5 performed best but couldn't complete the full end-to-end task (3D reconstruction is the bottleneck). An early empirical read on how far current models are from operating physical hardware autonomously.

---

Release Notes

Claude Code:

• v2.1.220 (Jul 25) — Bug fixes and reliability improvements
• v2.1.219 (Jul 24) — Adds claude-opus-5 as default Opus model (1M ctx, fast mode $10/$50/Mtok); nested subagents re-enabled at default depth 3; sandbox.network.strictAllowlist for deterministic host-denial in sandboxed commands; DirectoryAdded hook for mid-session working-directory registration; workflowSizeGuideline settable from any settings file; nested subagent forwarding in stream-json at depth 2+; MCP error details in mcp list; fixes dropped mid-stream answer on claude -p, stale Fable model billing label, and permission loss on self-hosted runner restart
• v2.1.218 (Jul 22) — /code-review now runs as background subagent; screen-reader announcements for deletions; fixes Windows path corruption of \u-prefixed segments (CJK character corruption bug); fixes left-arrow discard-with-no-undo; fixes multi-line paste collapsing via Ctrl+J; fixes /ultrareview failing on descriptive arguments
• v2.1.217 (Jul 21) — Subagents no longer spawn nested subagents by default (depth 5→1); concurrent subagent cap (default 20); emoji shortcode autocomplete (:heart: → ❤️); transcript-write failure warnings; fixes memory leak from truncated MCP tool outputs; fixes Windows auto-update leaving claude.exe missing; fixes background session isolation symlink canonicalization (security); fixes auto-compact for Opus 4.8 on Bedrock
• v2.1.216 (Jul 20) — sandbox.filesystem.disabled setting; fixes quadratic message-normalization slowdown in long sessions; fixes OAuth token expiry mid-session causing auto-mode denials; fixes AskUserQuestion neutral wording; fixes web session re-asking on idle timeout; fixes worktree-isolated subagent git redirect escapes

Agent SDKs (TypeScript):

• v0.3.220 (Jul 25) — Parity with Claude Code v2.1.220
• v0.3.219 (Jul 24) — Opt-in cancel_queued interrupt control; fast_mode_disabled_reason on result/init messages; DirectoryAdded lifecycle hook; sandbox.network.strictAllowlist and workflowSizeGuideline in settings types; fixes fast_mode_state reporting after model switch
• v0.3.218 (Jul 22) — SkillToolOutput.background: true for forked background skills; fixes api_error_status null for 429/529 mid-stream errors; canonicalModel and provider in modelUsage for accurate billing rate lookups
• v0.3.217 (Jul 21) — Subagents no longer spawn nested subagents by default (depth cap 5→1; set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH to restore deeper nesting); concurrent subagent cap added (default 20, override with CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS); fixes Remote Control pending permission prompts not re-sent to late-connecting clients
• v0.3.216 (Jul 20) — skippedLinks count on rewindFiles responses; tool_result_meta sidecar for classifying denied/interrupted/cancelled tool calls without string-matching; user_message_uuid and request_sent_wall_ms for cross-host latency correlation

Agent SDKs (Python):

• v0.2.128 (Jul 25) — Bundles Claude Code v2.1.220
• v0.2.127 (Jul 24) — Fixes premature stdin closure when background subagents are still running (previously caused SDK-MCP tool call failures and silent PreToolUse hook bypass); bundles v2.1.219
• v0.2.126 (Jul 22) — ResultMessage.terminal_reason surfaces loop-end cause ("aborted_streaming", "aborted_tools", etc.); ResultMessage.model_usage typed as dict[str, ModelUsage] with canonicalModel/provider fields; bundles v2.1.218
• v0.2.125 (Jul 21) — Bundles Claude Code v2.1.217
• v0.2.124 (Jul 20) — Security: blocks .bat/.cmd CLI spawning on Windows (BatBadBut/CVE-2024-27980 class); rejects cmd.exe metacharacters in resume/session_id on Windows; fixes dash-prefixed extra_args value binding; bundles v2.1.216

---

Action Items

Immediate:

• Review cloned repositories for unexpected symlinks before opening them in Claude Code — a newly disclosed flaw in the startup memory loader (verified on v2.1.215, no CVE or fixed version yet) can expose files outside the repo without approval. Watch the GitHub advisory feed for the patch.
• Update claude-agent-sdk (Python) to v0.2.124+ if you run Claude Code on Windows — it blocks .bat/.cmd CLI spawning vulnerable to cmd.exe metacharacter injection (BatBadBut class).
• Pin CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH if your workflows rely on nested subagents — the default depth moved from 5 to 1 to 3 within one week; riding the default means unpredictable orchestration behavior across versions.
• Audit any public-facing Claude Artifacts in your products — the SectopRAT campaign exploited user-generated Artifacts being served under claude.ai. Add explicit UI callouts where your product surfaces Artifact URLs.

By July 28 (tomorrow):

• Review your MCP server architecture for stateless compatibility — the final MCP 2026-07-28 spec publishes tomorrow, removing Sessions and initialize. Tier 1 SDKs preserve backwards compatibility, but check the MCP blog migration guides; TypeScript v2 splits the SDK into @modelcontextprotocol/server and @modelcontextprotocol/client.

Near-term:

• Try the built-in /claude-api migrate skill in Claude Code to move workloads to Opus 5 — it updates model strings and tunes prompts, and Opus 5's $5/$25 pricing makes the switch worth evaluating for anything currently on Opus 4.8 or Fable 5.
• Update billing code to use canonicalModel and provider from ModelUsage (TypeScript SDK v0.3.218+ / Python SDK v0.2.126+) rather than provider-specific model aliases, to ensure correct rate-table lookups across Bedrock, direct API, and future providers.

---

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.