TLDR;

Anthropic is extending Claude Fable 5 access on all paid plans and keeping Claude Code's weekly rate limits 50% higher through July 19 — announced Sunday, capping a week that also brought Claude Cowork to web and mobile. On the tooling side, Claude Code v2.1.207 makes auto mode generally available on Bedrock, Vertex AI, and Foundry — and those providers now default to Claude Opus 4.8. Six Claude Code releases (v2.1.202–v2.1.207) and seven Agent SDK releases piled on reliability fixes for background sessions, MCP timeouts, and interrupt semantics. Wiz researchers disclosed the GhostApproval symlink vulnerability affecting Claude Code and five other AI coding assistants — fixed in Claude Code v2.1.173 and later, update now. Anthropic also published the origin story of Claude Code, new interpretability research, and a hard reminder that Claude 3.5 Haiku on Vertex AI shut down July 5.

---

---

Headline

Claude Fable 5 on All Paid Plans, and Claude Code's Higher Rate Limits, Extended Through July 19

Announced July 12, 2026 — Anthropic is extending Claude Fable 5 access on all paid plans, and keeping Claude Code's weekly rate limits 50% higher, both through July 19. Fable 5 — Anthropic's most capable generally available model — returned globally on July 1, and these were the redeployment perks; they now run one more week. If you've been meaning to benchmark Fable 5 against Claude Opus 4.8 on your own workloads, or push a big Claude Code migration through while the ceiling is higher, this is the week to do it.

---

Product & Apps

Claude Cowork Expands to Web and Mobile

Announced July 7, 2026 — Cowork, Claude's agentic workspace previously exclusive to the desktop app, is rolling out to claude.ai on the web and to the iOS and Android apps, in beta on select paid plans. Chat and Cowork now share one home on web and desktop, with projects and artifacts living together across both; desktop remains the full experience, where Claude can also use local files and the browser. To mark the launch, Cowork usage limits are doubled through August 5.

---

API & Model Updates

API Keys Can Now Expire — Set It in the Claude Console

Shipped July 8, 2026 — You can now set an expiration when creating an API key or Admin API key in the Claude Console: a preset, a custom duration, or never. For keys with a lifetime of at least 7 days, Anthropic emails the creator before expiration. Existing keys are unaffected, and the Admin API reports each key's expiration in the expires_at field. Worth folding into your key-rotation policy — see Authentication.

⚠️ Model Deprecation — Claude 3.5 Haiku on Vertex AI: Already Shut Down

Claude 3.5 Haiku on Google Cloud Vertex AI was deprecated January 5, 2026 and permanently shut down July 5, 2026 — if you haven't migrated yet, calls to this model ID are already failing. See Google Cloud's partner model deprecations page for current model IDs and shutdown timelines including Claude 3 Haiku (August 23, 2026).

---

Claude Code & CLI

The Making of Claude Code

Published July 6, 2026 — Anthropic's detailed origin story of Claude Code — how it was built, what bets were made, and what the team learned. Worth reading for the architectural decisions and the reasoning behind the agent-first design; useful context when explaining Claude Code's capabilities to stakeholders or evaluating it against alternatives.

Claude Code v2.1.207: Auto Mode Generally Available on Bedrock/Vertex/Foundry; Opus 4.8 Becomes the Provider Default

Released July 11, 2026 — Auto mode no longer requires the CLAUDE_CODE_ENABLE_AUTO_MODE opt-in on Bedrock, Vertex AI, and Foundry (disable via disableAutoMode in settings — note it now reads from ~/.claude/settings.json, no longer from repo-resident .claude/settings.local.json). Bedrock, Vertex, and Claude Platform on AWS now default to Claude Opus 4.8. Also security-relevant: ${user_config.*} interpolation in shell-form plugin hook commands is now rejected (shell-injection fix) — plugins should use exec form or $CLAUDE_PLUGIN_OPTION_.

Also in this release: fixed terminal freezing and keystroke lag on long lists/tables/code blocks; fixed remote managed settings being silently recorded as consented in headless sessions; fixed spurious prompt-injection warnings from benign system updates; fixed auto-updater overwriting custom launcher scripts; fixed a crash loop in agent teams from malformed mailbox messages; fixed extensions.worktreeConfig polluting .git/config and breaking go-git tools like tea.

Claude Code v2.1.206: Push Remote Auto-Allow; Public Gateway /login; MCP Timeout Fix

Released July 10, 2026 — /commit-push-pr now auto-allows git push to the repo's configured push remote (not just origin). /doctor now proposes trimming checked-in CLAUDE.md files by stripping content Claude can derive from the codebase. Gateway: /login supports Anthropic-operated public gateway endpoints. Background agents now upgrade immediately after a Claude Code update rather than paying the slow stale-session upgrade cost on next attach. Fixed expired login showing a misleading model error instead of prompting /login; fixed --resume/--continue not responding to keyboard input; fixed per-server request_timeout_ms being ignored in fresh sessions (defaulting to 60s and timing out long MCP calls); fixed OAuth MCP servers requiring manual re-auth after a single failed token refresh.

Claude Code v2.1.205: Transcript-Tampering Rule; Windows Worktree Fix

Released July 8, 2026 — Added an auto mode rule blocking tampering with session transcript files. Fixed --json-schema silently producing unstructured output on invalid schemas. Fixed a message sent while Claude was working being silently lost at --max-turns limit. Fixed Windows worktree removal deleting files outside the worktree when NTFS junctions or symlinks were present. Fixed claude mcp add-from-claude-desktop getting stuck on server names with unsupported characters.

Claude Code v2.1.203: Background-Session Recovery; MCP roots/list Notifications

Released July 7, 2026 — Added a login-expiry warning so background sessions aren't interrupted mid-run. Added a grey ⏸ badge in the footer when in manual permission mode. Added the session's additional working directories to MCP roots/list with notifications/roots/list_changed on changes. Fixed background sessions becoming permanently unresponsive when the daemon's session token went stale (now auto-recovers). Fixed returning to claude agents stopping running subagents and re-running from scratch. Fixed background agents inheriting a stale PATH from the daemon (Windows). Fixed ANTHROPIC_BASE_URL being dropped in background sessions, sending API keys to the default endpoint.

Claude Code v2.1.202: Dynamic Workflow Size Setting; OTel Workflow Attributes

Released July 6, 2026 — New "Dynamic workflow size" setting in /config controls how many agents Claude generally spawns for dynamic workflows (small/medium/large — advisory, not enforced). Workflow-spawned agents now emit workflow.run_id and workflow.name OpenTelemetry attributes, enabling full workflow reconstruction from OTel data. Fixed Remote Control (mobile/web) commands failing with "Unknown command"; fixed images/files sent from Remote Control without a caption being silently dropped; fixed workflow scripts with unicode quote escapes being corrupted before parsing.

---

Agent SDK & Managed Agents

claude-agent-sdk-typescript v0.3.207: canUseTool 'allow' Without updatedInput Now Works

Released July 11, 2026 — Fixed canUseTool returning {behavior: 'allow'} without updatedInput being incorrectly rejected as a deny with a raw ZodError — the tool now runs with the original input per the documented contract. AgentToolCompletedOutput is now a published SDK type that matches the emitted object exactly.

claude-agent-sdk-typescript v0.3.206: command_lifecycle Frames; Accurate Zero-API Timing

Released July 10, 2026 — Added command_lifecycle frames to stream-json and SDK sessions, reporting each UUID-stamped message's terminal state (queued/started/completed/cancelled/discarded). Zero-API results no longer report stale duration_api_ms — useful for accurate billing attribution in multi-agent pipelines.

claude-agent-sdk-typescript v0.3.205: Typed Interrupt Receipts; still_queued UUIDs

Released July 8, 2026 — Interrupt control responses now include still_queued (UUIDs of queued async messages that will still run); Query.interrupt() returns a typed receipt; system/init advertises interrupt_receipt_v1 for feature detection. Added structured name and body fields to peer-message session events.

claude-agent-sdk-typescript v0.3.203: background_tasks_changed as Level; sdk.d.ts Typecheck Fix

Released July 7, 2026 — Added background_tasks_changed system message with the full set of live background tasks on every membership change — consumers can now track background activity as a level (current state snapshot) instead of pairing task_started/task_notification edge events. Also fixed stable releases shipping sdk.d.ts with unresolved type references that broke typechecking with skipLibCheck disabled.

claude-agent-sdk-typescript v0.3.202: parent_agent_id for Agent Tree Reconstruction

Released July 6, 2026 — Added parent_agent_id field to subagent session messages, enabling depth-2+ agent tree reconstruction from disk-persisted metadata. Fixed apply_flag_settings with a non-object value crashing the session instead of returning a control error.

claude-agent-sdk-python v0.2.116: Bundles Claude CLI v2.1.207; CI Workspace Trust Fix

Released July 11, 2026 — Bundles Claude CLI v2.1.207 — no separate claude install required, so CI pipelines using the Python SDK automatically receive CLI bug fixes on SDK upgrade. Also fixes CI workspace trust so Claude Code honors project-scoped permission grants in checkout directories. Install: pip install claude-agent-sdk==0.2.116. (v0.2.112–v0.2.115 were interim releases tracking CLI v2.1.203–v2.1.206.)

claude-agent-sdk-python v0.2.111: Zombie Subprocess Fix; NDJSON Whitespace Loss Fix

Released July 6, 2026 — Three substantive bug fixes: zombie CLI subprocess prevention (shielded subprocess cleanup from asyncio cancellation); silent whitespace loss on large NDJSON lines exceeding the 64 KiB stream buffer (could corrupt tool output or message content); uncaught TypeError when CLI emits a message with a plain-string content field. Also added a runtime warning when can_use_tool callback is registered alongside allowed_tools or bypassPermissions — the combination silently prevents the callback from firing.

---

Security

GhostApproval: Symlink Vulnerability Affecting Claude Code and Five Other AI Coding Assistants

Published July 9, 2026 — Wiz researchers disclosed "GhostApproval" (CWE-61), a symlink-following vulnerability that bypasses Human-in-the-Loop safety controls across Amazon Q Developer, Claude Code, Augment, Cursor, Google Antigravity, and Windsurf. In Claude Code's case, the symlink warning in the Edit/Write permission dialog shipped in v2.1.32 (February 5, 2026) prior to the report submission. Current versions (2.1.173+) resolve symlinks and warn before writing to sensitive files — update immediately if you're on an older build. AWS (CVE-2026-12958) and Cursor (CVE-2026-50549) have also patched; Augment and Windsurf had not shipped fixes at publication.

---

Enterprise & Business

UST Is Bringing Claude to Physical AI

Published July 9, 2026 — UST is integrating Claude into physical AI systems — relevant to builders working at the intersection of edge compute, robotics, or hardware-adjacent software. If you're building integrations that need to run closer to physical systems, this partnership signals Anthropic's direction in the embedded/physical AI space.

Government of Alberta Uses Claude for Cybersecurity Vulnerability Detection

Published July 6, 2026 — The Alberta government is using Claude to find and fix cybersecurity vulnerabilities across government systems — a significant public-sector deployment. Builders in the security tooling space should note this as a validated production use case for Claude in vulnerability scanning workflows.

Ben Bernanke Appointed to Anthropic's Long-Term Benefit Trust

Published July 9, 2026 — Former Federal Reserve Chair Ben Bernanke joins Anthropic's Long-Term Benefit Trust, which oversees the company's public benefit mission. This is a governance-layer appointment at the entity that holds Anthropic's public benefit mandate — context for enterprise customers evaluating Anthropic's long-term institutional stability.

---

Research & Safety

An Off Switch for Dual-Use Knowledge in AI Models

Published July 8, 2026 — Anthropic research on selectively suppressing dual-use knowledge in AI models without degrading general capability. Directly relevant to builders in regulated industries or those building on top of Claude for applications with compliance requirements around information access — this signals the research direction behind future policy controls.

A Global Workspace in Language Models

Published July 6, 2026 — Interpretability research identifying a global workspace mechanism in language models — a bottleneck through which distributed computations converge. Useful background for builders designing multi-agent systems or working on steering/intervention approaches.

Introducing a Way to Reflect on How You Use Claude

Published July 9, 2026 — Anthropic is launching a reflection feature for Claude.ai users to review their usage patterns. Consumer-facing but worth tracking — shifts in how Anthropic frames user relationship with Claude often precede API-level policy updates.

---

Action Items

Immediate — already broken:

• Claude 3.5 Haiku on Vertex AI is already shut down (July 5, 2026). If calls are failing, migrate to a current model ID — see Google Cloud's partner model deprecations page for replacements.
• GhostApproval symlink vulnerability — update to Claude Code v2.1.207 (v2.1.173 or later contains the fix). If Augment or Windsurf is in your stack, those vendors had not shipped fixes at disclosure (July 9).

By July 19, 2026:

• Fable 5 promo window closes — all paid plans have Claude Fable 5 access, and Claude Code's weekly rate limits stay 50% higher, through July 19. Run your heaviest jobs and model comparisons this week.

By July 28, 2026:

• MCP 2026-07-28 spec goes final — test your MCP servers against the beta SDKs now. Key breaking changes: stateless core (no Mcp-Session-Id, no initialize handshake); missing-resource errors move from MCP-custom -32002 to JSON-RPC -32602; Roots, Sampling, and Logging deprecated (annotation-only, keep working through at least mid-2027). Nothing breaks automatically on July 28, but unvalidated servers will drift from spec.

By August 5, 2026:

• Migrate off Claude Opus 4.1 before the hard retirement date (August 5). Recommended replacement: claude-opus-4-8.
• Audit temperature/top_p/top_k usage in any integration targeting Claude Opus 4.7+, Claude Opus 4.8, or Claude Sonnet 5 — setting these to non-default values now returns a 400 error.

By August 23, 2026:

• Claude 3 Haiku on Vertex AI is deprecated and available to existing customers only; migration required before the August 23 shutdown. See Google Cloud 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.