TLDR;

Fable 5 is back: the United States government lifted the June 12 export controls on June 30, Claude Fable 5 rolled out globally on July 1, and Mythos 5 is restored for approved United States organizations. Claude Sonnet 5 is now the default model in Claude Code — with two breaking API changes and a new tokenizer to check before you migrate. Claude Code shipped six releases (v2.1.196–v2.1.201), beta TypeScript and Python SDKs landed for the stateless MCP spec finalizing July 28, and three deadlines are coming fast: individual usage analytics defaults on July 11, Opus 4.7 fast mode is removed July 24, and Opus 4.1 retires August 5.

---

---

API & Model Updates

Fable 5 Is Back: Export Controls Lifted, Global Rollout, Mythos 5 Restored

June 30, 2026 (updated July 1) — The United States government lifted the June 12 export-control directive on June 30, and Claude Fable 5 became available globally on July 1 across the Claude Platform, claude.ai, Claude Code, and Claude Cowork. Claude Mythos 5 is restored for approved United States organizations (approval granted June 26). The redeployment ships with a new safety classifier targeting the jailbreak method behind the original suspension, and Anthropic proposed a joint jailbreak-severity framework with Amazon, Microsoft, and Google. Mind the subscription window: on Pro, Max, Team, and select Enterprise plans — including Claude Code subscription usage — Fable 5 is included for up to 50% of weekly usage limits only through July 7, after which it switches to usage credits only. If you rerouted Fable 5 workloads to Claude Opus 4.8 during the suspension (as we suggested while claude-fable-5 was dark), it's time to re-evaluate.

Claude Sonnet 5: Now the Default in Claude Code and on claude.ai Free and Pro

Released June 30, 2026 — Claude Sonnet 5 is the new default model in Claude Code (v2.1.197+) and for Free and Pro plans on claude.ai, with a 1M-token context window at introductory pricing of $2/$10 per million tokens through August 31 (then $3/$15). Before migrating, check three things. Two breaking API changes: manual extended thinking (thinking: {type: "enabled", budget_tokens: N}) is removed and returns a 400 error, and setting sampling parameters (temperature, top_p, top_k) to non-default values also returns a 400. One default change: adaptive thinking is on by default — turn it off with thinking: {type: "disabled"} if you need the old behavior. Sonnet 5 also ships a new tokenizer that produces approximately 30% more tokens for the same text (Anthropic's own figure; roughly 1.0–1.35x depending on content type) — per-token prices are unchanged, but token counts, context budgets, and anything you meter in tokens will shift, so re-run your token counting.

---

Claude Code & CLI

v2.1.197: Claude Sonnet 5 Ships as Default Model

June 30, 2026 — The headline release of the week: Claude Sonnet 5 is now the default model in Claude Code with a 1M-token context window and the promotional $2/$10 per million token pricing through August 31. Update to v2.1.197 or later (npm i -g @anthropic-ai/claude-code) to get access; see the migration notes in the API section above.

v2.1.198: Subagents Run in the Background by Default; Claude in Chrome Generally Available

July 1, 2026 — Subagents now run in the background by default, so Claude keeps working while they execute (previously a gradual rollout). For background agents launched from claude agents, sessions that need input or finish now fire the Notification hook with agent_needs_input / agent_completed identifiers — and they now commit, push, and open a draft PR when they finish code work in a worktree instead of stopping to ask. Claude in Chrome is now generally available. The built-in Explore agent inherits the main session's model (capped at Opus) instead of running on Haiku, and subagents and context compaction both inherit the session's extended thinking configuration. Also added: a /dataviz skill for chart and dashboard design with a runnable color-palette validator, and Claude Platform on AWS (anthropicAws) as a Gateway upstream provider.

v2.1.199: Stacked Skill Invocations + SSL/Streaming Fixes

July 2, 2026 — Stacked slash-skill invocations like /skill-a /skill-b do XYZ now correctly load all leading skills (up to 5), not just the first. Critical reliability fixes: SSL certificate errors (TLS-inspecting proxies, missing NODE_EXTRA_CA_CERTS, expired certificates) now fail immediately with an actionable fix hint instead of burning retries; streaming responses interrupted by a mid-stream server error now preserve the partial output with an incomplete-response notice; and subagents cut off by rate limits or server errors now return their partial work to the parent instead of silently failing.

v2.1.200: Permission Mode Now Labeled "Manual"; AskUserQuestion No Longer Auto-Continues

July 3, 2026 — Two behavior changes worth knowing: AskUserQuestion dialogs no longer auto-continue by default (opt into an idle timeout via /config), and the "default" permission mode is now labeled "Manual" across the CLI, --help, VS Code, and JetBrains (--permission-mode manual and "defaultMode": "manual" are accepted alongside "default"). Major background-agent reliability fixes: a stale daemon.lock whose process ID the operating system reused no longer blocks agents after a crash, reinstalled older builds can no longer take over the daemon, subagents that hit rate limits before producing any output now fail cleanly, and roster corruption issues are resolved. Also fixed: a startup crash when disabledMcpServers or enabledMcpServers in .claude.json is set to a non-array value.

v2.1.201: Claude Sonnet 5 System Role Change

July 3, 2026 — Claude Sonnet 5 sessions no longer use the mid-conversation system role for harness reminders. A small one, but it lands the same day as v2.1.200 — the sixth release of the week.

v2.1.196: Org Default Models, Clickable File Attachments, MCP Security Fix

June 29, 2026 — Admins can now set organization default models in the org console; it shows as "Org default" or "Role default" in /model when users haven't picked one. Sessions now get readable default names at start. Clickable file attachments in chat: Cmd/Ctrl-click reveals the file in Finder/Explorer. Security fix: claude mcp list/get no longer spawns .mcp.json servers that a repo self-approved via a committed .claude/settings.json — untrusted workspaces now show ⏸ Pending approval.

---

Agent SDK & Managed Agents

claude-agent-sdk-typescript v0.3.199: requestId for canUseTool, Blocked Field, Masked Credentials

July 2, 2026 — Three new capabilities for SDK-hosted permission flows: requestId added to canUseTool callback options for correlating out-of-band permission responses; returning null from canUseTool suppresses the SDK's automatic control response; a blocked field on workflow_agent progress events indicates when an agent was blocked by the auto-mode safety classifier; and mode:"mask" plus per-credential injectHosts land in the sandbox.credentials settings types for injecting masked credentials into sandboxed commands.

claude-agent-sdk-typescript v0.3.200: 'manual' Permission Mode Alias; set_model Validation Fix

July 3, 2026 — 'manual' is now accepted as an alias for 'default' in SDK permission-mode inputs, matching the Claude Code v2.1.200 rename. onSetPermissionMode now fires correctly for SDK-hosted Remote Control sessions, and set_model control requests reject unrecognized model strings before latching — previously invalid models would silently latch.

claude-agent-sdk-typescript v0.3.198: canUseTool Warning, per-server request_timeout_ms, Bug Fixes

July 1, 2026 — A runtime warning now fires when canUseTool is configured alongside allowedTools or bypassPermissions, which shadow the callback. Per-server request_timeout_ms option added to the mcp_set_servers control request. Fixed SDKUserMessage.isSynthetic not being mapped to isMeta, which could cause synthetic messages to be treated as real user messages, and fixed workflow progress events silently dropping the earliest agents from the list while the phase counter stayed correct.

claude-agent-sdk-typescript v0.3.196: prompt_id in Hook Payloads, Dedup Fix

June 29, 2026 — prompt_id field added to hook input payloads for correlating hook events with OpenTelemetry prompt-level events. Fixed control protocol deduplication dropping tool-use IDs after 1,000 resolutions, which could cause duplicate tool_result deliveries in long-running sessions. (v0.3.197 and v0.3.201 also shipped this week as parity-only releases tracking Claude Code.)

Managed Agents: Live-Preview Event Deltas on Session Streams

June 30, 2026 — Managed Agents session streams gain event_deltas[] live-preview events (event_start / event_delta), letting harnesses render agent output as it's produced instead of waiting for completed events.

---

MCP Ecosystem

Beta TypeScript and Python SDKs Ship for the Stateless 2026-07-28 MCP Spec

June 29, 2026 — Beta TypeScript and Python SDKs are out implementing the 2026-07-28 release candidate (locked May 21) — the largest MCP revision since launch, and your window to test before the final specification ships July 28, 2026. What changes: a stateless core — SEP-2575 removes the initialize/initialized handshake and SEP-2567 removes the Mcp-Session-Id header, with protocol version, client info, and capabilities traveling in _meta on every request — so a remote server that needed sticky sessions and a shared session store can now run behind a plain round-robin load balancer. Two official extensions arrive (server-rendered UIs via MCP Apps, long-running work via Tasks), authorization aligns more closely with how OAuth 2.0 and OpenID Connect are deployed in practice, and three core primitives — Roots, Sampling, and Logging — are deprecated (annotation-only: they keep working in this release and every spec version published within a year of it, so through at least mid-2027).

---

Research & Safety

Fable 5 Safeguards Deep Dive + a Proposed Industry Jailbreak-Severity Scale

July 2, 2026 — The companion piece to the redeployment: how Fable 5's cyber classifiers work, plus a proposed five-band jailbreak-severity scale (CJS-0 through CJS-4) developed with the Project Glasswing partners so labs grade jailbreak reports on a shared scale. Anthropic also opened a HackerOne bounty specifically for Fable 5 cyber jailbreaks.

Claude Science: An AI Workbench for Researchers, with up to $30K in Credits

June 30, 2026 — Claude Science is a new AI workbench app for researchers, in beta on macOS and Linux for Pro, Max, Team, and Enterprise plans. It connects to 60+ scientific databases, reaches high-performance computing over SSH and Modal, and works artifact-first with an auditable history. Anthropic is offering up to $30,000 in credits per research project — applications are open through July 15.

---

Competitor Dev Tools

Cursor for iOS: Native App in Public Beta on All Paid Plans

June 29, 2026 — Cursor launched a native iOS app in public beta for all paid plans: pick a repo and launch a cloud agent from your phone, get push notifications (plus lock-screen Live Activities) when an agent finishes or needs input, review diffs, and merge the PR directly from the app. Sessions can move from local to cloud to keep running with your laptop closed.

Cursor: Team MCPs and Organization Groups in Team Marketplaces

June 30, 2026 — Admins can now configure Team MCP servers once and distribute them across cloud agents, the agents window, IDE, and CLI via the team marketplace. Team marketplaces also gain organization-group access controls alongside the existing team-level SCIM directory groups.

GitHub Copilot: Browser Tools Generally Available, Agent Session Streaming in Preview

July 1–2, 2026 — A dense two days for Copilot: browser tools in VS Code are now generally available, letting the agent drive a browser during tasks; agent session streaming entered public preview; Kimi K2.7 Code joined the model picker; and the CLI and SDK gained AI credit session limits for capping cost per agent session, alongside auto model selection that routes by task.

Windsurf's Cascade Agent Reaches End of Support; Devin Local Is the Successor

Support window closed July 1, 2026 — Per Cognition's June 2 announcement, the legacy Cascade agent's support window ended this week: Windsurf is now Devin Desktop, and Devin Local — a from-scratch Rust rewrite Cognition says is up to 30% more token-efficient, with subagent support — is the successor local agent (still labeled preview, and not supported in the JetBrains plugin). If your team still runs Cascade workflows or memories, the migration path is Devin Local's skills system.

---

Enterprise & Business

Claude Apps Gateway: A Self-Hosted Control Plane for Claude Code on Bedrock and Google Cloud

June 29, 2026 — Running Claude Code on Amazon Bedrock or Google Cloud used to mean a cloud credential per developer and no per-developer spend visibility. The new gateway is a self-hosted control plane (a single stateless container backed by PostgreSQL) that adds corporate single sign-on via OpenID Connect, centrally enforced managed settings, per-user cost attribution over OpenTelemetry to a collector you operate, routing across the Claude API, Bedrock, and Google Cloud with optional failover, and daily/weekly/monthly spend caps per organization, group, or user. Anthropic is publishing the gateway protocol so other gateway developers can implement the same features. Available now.

Claude in Microsoft Foundry Is Now Generally Available

June 29, 2026 — Claude models are generally available in Microsoft Foundry, hosted on Azure with the authentication, billing, and governance controls Azure teams already use — including a US data zone for data-residency requirements. Claude Opus 4.8 and Claude Haiku 4.5 are available in the Messages API to start, with prompt caching and extended thinking; usage lands on a single Azure invoice and draws down Microsoft Azure commitments for eligible Enterprise Agreement customers. The previous Foundry Preview continues as the "hosted on Anthropic" option for the full API feature set.

New Claude Admin Analytics: Spend Dashboards, Alerts, and Claude Code Usage/Value Tabs — Individual Analytics On by Default July 11

July 2, 2026 — Anthropic shipped richer admin analytics and spend controls: the dashboard now shows usage and cost by group and by user (filterable by SCIM groups), with spend-threshold alerts at 75% and 90% of an org-level limit. The Claude Code admin console gains two tabs — Usage (active developers, session counts, top commands, updated daily) and Value (estimated productivity lift, cost per commit, and annual value, with every formula visible and every input adjustable). Important for admins: individual usage analytics is currently off by default but will be on by default starting July 11, 2026 — if you want to keep member-level visibility off, change the setting before that date.

---

Action Items

Immediate:

• Fable 5 is back globally — if you rerouted claude-fable-5 workloads to Opus 4.8 during the June 12 suspension, re-evaluate now that access is restored.
• Tomorrow, July 7, is the last day Fable 5 is included in Pro, Max, Team, and select Enterprise subscriptions (up to 50% of weekly usage limits, Claude Code included) — after that it's usage credits only, so enable credits if your team needs continued access.
• Migrating to Claude Sonnet 5 (now the Claude Code and claude.ai Free/Pro default)? Remove thinking: {type: "enabled", budget_tokens: N} and any non-default temperature/top_p/top_k first — both now return 400 errors — and re-run token counting: the new tokenizer produces ~30% more tokens for the same text.
• If you use AskUserQuestion dialogs in automation, note the v2.1.200 change — dialogs no longer auto-continue by default; opt into an idle timeout via /config.
• Audit any Claude Code setups that rely on .mcp.json servers self-approved via a committed .claude/settings.json — v2.1.196 blocked this flow; untrusted workspaces now require manual approval.
• Individual usage analytics in Claude Enterprise will be on by default July 11, 2026 — if you want member-level visibility off, change the setting before then.

By July 15, 2026:

• Researchers: Claude Science credit applications (up to $30,000 per research project) close July 15.

By July 24, 2026:

• Migrate any calls to claude-opus-4-7 with speed: "fast" to Claude Opus 4.8 fast mode (research preview) — Opus 4.7 fast mode is removed July 24 and requests will hard-error, with no fallback to standard speed. (Opus 4.6 fast mode was already removed June 29 with the opposite behavior: those requests silently run at standard speed and standard billing.)

By late July:

• Review the 2026-07-28 MCP release candidate and test against the new beta SDKs if you run remote MCP servers — the stateless changes (no Mcp-Session-Id, no initialize handshake) finalize July 28. Roots, Sampling, and Logging are deprecated but keep working until at least mid-2027.
• If you build on GitHub Models, plan your migration — it is fully retired on July 30, 2026.

By August 5, 2026:

• Migrate off claude-opus-4-1-20250805 — the Claude API retires this model on August 5; Anthropic recommends Claude Opus 4.8.

---

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.