The week's story is the most powerful model Anthropic ever shipped getting pulled less than 72 hours after launch. On June 9, Fable 5 and Mythos 5 went live as the first Mythos-class models made safe for general use -- Fable at $10 / $50 per million tokens, Mythos 5 through Project Glasswing partners only. On June 12 at 5:21 p.m. ET, Commerce Secretary Lutnick sent Dario Amodei an export control directive citing national security concerns over a possible jailbreak, ordering the suspension of all foreign-national access to both models -- anywhere in the world, including foreign-national Anthropic employees. Anthropic complied by midnight but publicly disputed the government's underlying assessment. All other Claude models -- Opus 4.8, Sonnet, Haiku -- keep running unchanged, and the company says it is "working to restore access as soon as possible."

That is the front-door story. The back-door story is that Claude Code's release train did not stop for any of it -- seven versions in seven days, headlined by v2.1.172 letting sub-agents spawn their own sub-agents (up to five levels deep) and v2.1.170 wiring Fable 5 into the CLI on day one. Apple used WWDC to put Claude, Gemini, and GPT inside Xcode 27 as first-class agents. And we shipped vew, our desktop browser with an agent built in: twenty-six point releases in eight days, a first public launch, and a position on the agent-browser problem that runs in the opposite direction of the cloud-brokered headless setups everyone else is shipping.

From us

  • vew 0.2.9 through 0.4.26: a real browser with an agent built in - first public cut shipped June 2-10 across Windows (x64 and ARM64), macOS, and Linux. The agent reads the live DOM you are looking at and drives the browser -- click, type, navigate, fill forms -- on your real local session, with human-in-the-loop confirmation for purchases and other irreversible steps. A built-in @yawlabs/vew-mcp bridge (off by default, 127.0.0.1:9222 only) lets Claude Code or any MCP client drive this Vew instead of launching its own headless Chromium. Privacy posture is the whole pitch: no Google search defaults, card and SSN numbers Luhn-stripped before they reach the model, sensitive pages gated, and your logged-in sessions never leave the machine

Anthropic this week

  • Statement on the US government directive to suspend access to Fable 5 and Mythos 5 - June 12 export control order from Commerce citing a possible jailbreak, applied to all foreign nationals worldwide; Anthropic complied and publicly disputed the assessment. Restoration timeline is open-ended; all other Claude models and tiers are unaffected
  • Claude Fable 5 and Claude Mythos 5 launch (June 9) - first Mythos-class models made safe for general use; Fable at $10 / $50 per million tokens, Mythos 5 through Project Glasswing cyberdefender partners; new safety classifiers on cyber, biology/chemistry, and distillation with automatic fallback to Opus 4.8; 30-day data retention on Mythos-class traffic for safety analysis only

Claude Code this week

  • v2.1.170 - June 9: introduces Claude Fable 5 in the CLI on launch day; fixes VS Code integrated-terminal sessions not saving transcripts (and not appearing in --resume) when shell-inherited env vars confused the harness
  • v2.1.172 - June 10: sub-agents can now spawn their own sub-agents, up to five levels deep; Bedrock now reads region from ~/.aws/config when AWS_REGION is unset; /plugin marketplace gets a search bar; WebFetch(domain:*.example.com) wildcards now match subdomains; a stack of background-session and availableModels allowlist fixes
  • v2.1.173 - June 11: Fable 5 [1m] suffix normalization; spurious Windows "sandbox dependencies missing" warning fixed
  • v2.1.174 - June 12: /model picker stops hiding the family that Default resolves to (Opus, Sonnet rows now visible on the right plan); Bedrock GovCloud region prefix fix; background sessions stop inheriting another session's ANTHROPIC_* provider env (gateway URL, custom headers, /model aliases); VS Code /usage gains per-skill / per-agent / per-plugin / per-MCP attribution over 24h and 7d
  • v2.1.175 - June 12: enforceAvailableModels managed setting -- when on, a Default that would resolve to a disallowed model falls back to the first allowed one, and user or project settings can no longer widen a managed availableModels list
  • v2.1.176 - June 12: session titles now generated in the conversation's language; footerLinksRegexes setting for regex-matched link badges; Bedrock awsCredentialExport caches credentials until their Expiration; availableModels alias picks can no longer be redirected to a blocked model via ANTHROPIC_DEFAULT_*_MODEL; Auto mode classifier falls back to the best available Opus when Fable 5 is enabled for the org but Opus 4.8 is not; large batch of Remote Control, background-session, and tmux-over-SSH fixes
  • v2.1.177 - June 13: changelog-only release

The broader week

  • Apple ships Xcode 27 with native agentic coding from Anthropic, Google, and OpenAI - WWDC 2026 (June 8) puts Claude, Gemini, and GPT inside the IDE as first-class agents that can write code, run tests, interact with the simulator, and validate their own output. Not plugins, not third-party extensions -- the agents are built into Xcode itself, paired with a Device Hub that lets them exercise real hardware. Xcode's installer also shrinks by 30% and requires Apple silicon
  • OpenAI expands Codex with role-specific plugins, Codex Sites, and ChatGPT integration - announced June 2, rolling out over the following weeks. Six role-specific plugins bundle 62 business apps and 110 automated skills for non-developer workflows; Codex Sites is a preview of shareable, hosted apps and pages built from natural language; annotations let users revise specific parts of documents, slides, and sheets. Codex is also landing inside ChatGPT as a unified agent surface