Fable 5 launched, then the US government suspended it

TLDR;

Anthropic launched Claude Fable 5 — its first publicly available Mythos-class model — on Tuesday (June 9), then on Friday (June 12) a US government export-control directive forced Anthropic to suspend Fable 5 and Mythos 5 for all users; access is still down with no restoration date. Meanwhile, two deadlines land today (June 15): Claude Sonnet 4 and Opus 4 API calls start erroring, and the agent/headless billing split goes live. Claude Code shipped eight releases this week (v2.1.169–v2.1.176) — nested subagents, Bedrock GovCloud fixes, and enforceAvailableModels for enterprise lockdown — and dynamic workflows reached general availability.

---

---

API & Model Updates

Claude Fable 5 and Mythos 5 Launch — First Publicly Available Mythos-Class Model

Released Tuesday, June 9, Claude Fable 5 is Anthropic's first generally available Mythos-class model, shipped day-one on the Claude API, Amazon Bedrock, Vertex AI, Microsoft Foundry, and GitHub Copilot at $10 per million input tokens and $50 per million output tokens (double Opus 4.8, with the 90% prompt-caching discount intact). Mythos 5 is the same underlying model with cyber safeguards lifted, restricted to Project Glasswing partners. Anthropic says Fable 5 leads on agentic coding (80.3% on SWE-Bench Pro versus Opus 4.8's 69.2%) and that Stripe used it to run a codebase-wide migration on a 50-million-line Ruby codebase in a day — work Anthropic estimates would have taken a team over two months by hand. Note: the model is suspended as of June 12 (see below), so day-one availability is currently moot.

---

🚨 US Government Export-Control Directive Suspends Fable 5 and Mythos 5 — Access Down for All Users

On Friday, June 12 (the directive arrived at 5:21pm ET), the US government — citing national-security export-control authority — ordered Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign national, whether inside or outside the United States, including Anthropic's own foreign-national employees. Because Anthropic cannot filter access by nationality in real time, it disabled both models for all customers worldwide to comply; every other Claude model (Opus 4.8, Sonnet 4.6) is unaffected. New sessions fall back to your default model or Opus 4.8, and API calls to claude-fable-5 return errors.

Anthropic is complying under protest. Per its statement, the government's letter gave no specific details; Anthropic's understanding is that the concern is a method of "jailbreaking" Fable 5, but it says the findings it was shown are minor or benign and that comparable capability is widely available from other public models (it cited OpenAI's GPT-5.5). Axios reported that Commerce Secretary Howard Lutnick sent the directive letter to Dario Amodei. As of this issue (June 15), access has not been restored and there is no ETA.

---

Claude Sonnet 4 and Opus 4 Retire Today (June 15) — API Calls Will Error

claude-sonnet-4-20250514 and claude-opus-4-20250514 — and the aliases claude-sonnet-4-0 / claude-opus-4-0 — stop accepting requests on June 15, 2026 (9am PT). After that, calls to these IDs return errors with no automatic fallback. Anthropic's recommended successors are claude-sonnet-4-6 and claude-opus-4-8. A global search for 20250514 across your code is the fastest way to find every call that needs updating.

---

Agent SDK / claude -p Billing Splits from Subscription Today (June 15)

Starting June 15, the Claude Agent SDK, claude -p headless mode, Claude Code GitHub Actions, and third-party agents move off your plan's subscription limit onto a separate monthly credit pool ($20 Pro / $100 Max 5x / $200 Max 20x), metered at full API rates with no rollover. Interactive Claude Code in the terminal or IDE is explicitly unaffected and keeps drawing from your subscription.

---

What Changes for Your Code When Fable 5 Returns: Refusals, Tokenizer, Retention

While Fable 5 is suspended, three behaviors are worth knowing before you wire it back in. (1) Safety classifiers return stop_reason: "refusal" on a successful HTTP 200 — not a network error, and not billed at Fable rates — with the server retrying on Opus 4.8; the Python, TypeScript, Go, Java, and C# SDKs ship refusal-fallback middleware, but anything calling the API directly must handle this stop reason explicitly. (2) The Opus-4.7-era tokenizer produces ~30% more tokens for the same text — recalibrate any cost or rate-limit math using the token-counting API with model: "claude-fable-5". (3) Mandatory 30-day data retention, with no zero-data-retention option — both Fable 5 and Mythos 5 are Covered Models; Microsoft restricted internal employee use over this policy while still selling the model to Copilot/Foundry customers.

---

Claude Code & CLI

Claude Code v2.1.170 — Nested Subagents (up to depth 5)

The June 9 release (which also added the now-suspended Fable 5 access) introduced nested subagent support: subagents can spawn their own subagents up to five levels deep, enabling deeper autonomous task decomposition.

---

Claude Code v2.1.172–v2.1.176 — Bedrock, Model Picker, and Hook Fixes

A dense run of six releases (June 10–12):

• Bedrock: reads the AWS region from ~/.aws config when AWS_REGION is unset; fixes GovCloud regions (us-gov-*) sending the wrong inference-profile prefix and causing 400 errors; caches awsCredentialExport credentials until their real Expiration instead of a fixed one hour.
• Enterprise lockdown: the new enforceAvailableModels managed setting makes the availableModels allowlist also constrain the Default model and blocks user/project settings from widening a managed allowlist.
• Fixes: sessions stuck on 1M context without usage credits now auto-compact; the /model picker shows Opus as its own row on Max/Team Premium/Enterprise; skill hot-reload no longer resends the full skill list on single-skill changes; hook if conditions for Edit(src/**), Read(~/.ssh/**), and Read(.env) now match correctly; Remote Control no longer silently switches the session model on web/mobile connect; tmux clipboard copy fixed.
• Heads-up: v2.1.171 was skipped on npm — pin to 2.1.172 or later in CI. /plugin gained a search bar, and session titles now generate in your conversation's language (pin with the language setting).

Run claude update to pull these in. (v2.1.176 release notes)

---

Claude Code v2.1.169 — --safe-mode, /cd Command, disableBundledSkills

The June 8 release adds --safe-mode (or CLAUDE_CODE_SAFE_MODE) to start with all customizations disabled for troubleshooting; /cd moves a session to a new working directory without breaking the prompt cache; and disableBundledSkills hides bundled skills and built-in slash commands from the model. Also fixes enterprise managed MCP policies not being enforced on reconnect.

---

Dynamic Workflows in Claude Code — Now Generally Available

Dynamic workflows reached general availability on June 10. Describe a complex task and Claude writes its own orchestration script, fans out parallel subagents, and verifies results before surfacing them (trigger with ultracode or by asking Claude to "create a workflow"). Two headline migrations: Jarred Sumner ported the Bun runtime from Zig to Rust — roughly 750,000 lines, 99.8% test-suite pass rate, 11 days from first commit to merge, presented at Code with Claude Tokyo — and Anthropic says Stripe used Fable 5 plus workflows for the 50-million-line Ruby migration noted above.

---

MCP Connector Observability Dashboard for Developers

Launched June 8: connector developers now have an observability dashboard showing active users, tool-call counts, directory ranking, per-tool error rates and latency, and usage broken down by Claude product (Claude, Claude Code, Cowork, etc.). A new in-app connector-submission portal is also live.

---

Claude Support for Apple's Foundation Models Framework

Released June 8: a new Swift package lets Apple developers use Apple's Foundation Models framework to call Claude for multi-step reasoning, code generation, and long-context tasks — @Generable-annotated typed Swift values arrive at the Claude API as structured data and responses stream back into SwiftUI views. At WWDC 2026 (June 9), Apple announced it would open-source the framework and open it to third-party models including Claude and Gemini.

---

Agent SDK & Managed Agents

Claude Agent SDK (TypeScript) v0.3.169–v0.3.176 — Fable Types, Fallback Triggers, Resume Fixes

Across four June releases: adds the claude-fable-5 model and fable alias to SDK types (v0.3.170); system/model_fallback now fires for all triggers — overloaded, server_error, and last_resort join model_not_found and permission_denied, so update any code that switches on the trigger field (v0.3.174); the plugins option accepts skipMcpDiscovery: true per plugin (v0.3.172); and v0.3.176 fixes turn result messages dropped when multiple turns complete during a background agent or workflow, plus background/remote/MCP task state not restored on session resume. Also: an experimental usage API and an sse browser option (v0.3.169). The Python SDK v0.2.96 pins mcp below 2.0.0 — check for conflicts if you manage that version yourself.

---

Claude Managed Agents — Scheduled Deployments and Vault Environment Variables (Public Beta)

Launched June 9 at Code with Claude Tokyo: agents can now run on a cron schedule, and securely reach CLI tools via vault-stored environment variables injected at the network boundary (the agent never sees your actual key). New self-hosted sandbox guides cover five providers — Blaxel, E2B, Google Cloud, Namespace Labs, and Superserve.

---

Developer Tools & Community

Using Fable 5 to Edit Its Own Launch Video

Thariq (Anthropic) wrote up editing the Fable 5 launch video end-to-end with Fable 5 plus Claude Code — the model wrote the code and tool calls for transcription, ffmpeg, color grading, Figma MCP, and Remotion rendering. The slide deck is public if you want to replicate the workflow.

---

Paving the Way for Agents in Biology

Anthropic's June 8 research post examines why AI has advanced faster in coding than biology ("bio databases are like cities built before cars") and what infrastructure changes would make them agent-navigable. Relevant for builders working on biomedical applications.

---

Enterprise & Business

Anthropic Partners with TCS and DXC for Regulated Industries

Two systems-integrator alliances landed this week: Tata Consultancy Services (June 12) and DXC Technology (June 11) will bring Claude into banking, aviation, healthcare, and other compliance-heavy enterprise systems. Relevant if you sell into regulated verticals.

---

First Anthropic Public Record Released

Published June 12, Anthropic's first public-transparency report covers model behavior, safety evaluations, and policy commitments.

---

Research & Safety

Dario Amodei on the AI Exponential, plus Claude Corps

Anthropic's CEO published "Policy on the AI Exponential" (June 10), arguing AI is advancing faster than policymaking institutions can handle, alongside three initiatives — including Claude Corps (June 11), a $150M fellowship matching 1,000 early-career people with US nonprofits to use Claude in advancing their missions. Worth reading for signal on where Anthropic's regulatory posture is heading.

---

Action Items

Today (June 15):

• Migrate off claude-sonnet-4-0 and claude-opus-4-0 (and the -20250514 IDs) — API calls return errors after today. Move to claude-sonnet-4-6 and claude-opus-4-8; search your code for 20250514.
• Audit headless agent billing — the Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party agents move to a separate credit pool today. Verify you have credits and that cost monitoring is updated; interactive Claude Code in the terminal is unaffected.
• Update Claude Code to v2.1.176 — run claude update for the hook-path, Bedrock GovCloud, and Remote Control model-switch fixes.

By June 18, 2026:

• Gemini CLI free tier shuts down — if any tooling relies on Gemini Code Assist for individuals, Gemini CLI free tier, or Google AI Pro/Ultra IDE extensions, they stop serving on June 18. Migrate to a paid key or Antigravity CLI.

By July 1, 2026:

• Windsurf / Devin Desktop: remove Cascade references — Cascade is end-of-life July 1; update any workflow or script that names it to use Devin Local.

By August 5, 2026:

• Migrate off Claude Opus 4.1 — deprecated June 5, retires on the Claude API August 5. Anthropic recommends Claude Opus 4.8 as the replacement.

---

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.