TL;DR: On June 12, 2026, a US government export-control directive forced Anthropic to disable Claude Fable 5 and Mythos 5 for every user on earth — and as of today, June 25, they’re still dark with no restoration date. The lesson for developers isn’t about one model; it’s that any cloud model can vanish overnight with zero notice. If your editor is hard-wired to a single model, you’re one directive away from a dead workflow. Build a fallback now.

	Cursor (fallback)	Cline / Claude Code	OpenCode + local Ollama
Failover speed	Instant model switch (chat/plan only)	Swap profile, ~30 sec	Already local — never goes down
Agentic work covered	No — agent locks to Cursor’s backend	Yes, full agent	Yes, full agent
Cost when primary dies	Same plan	Whatever the new model bills	$0
The catch	Composer/autocomplete can’t use external models	You manage API keys	Slower; needs a GPU

Honest take: Keep using whatever model is fastest today, but architect your stack so swapping models is a 30-second config change, not an afternoon. The developers who shrugged off June 12 all had one thing in common — a model-agnostic agent (Cline, Claude Code, or OpenCode) plus a local Ollama fallback that no government order can switch off.

What actually happened

Anthropic launched Claude Fable 5 to the public on June 9, 2026 — a Mythos-class model with exceptional software-engineering scores, briefly the best coding backend you could point Cursor or Cline at. Three days later it was gone.

At 5:21 p.m. ET on June 12, Anthropic received a legally binding US government export-control directive. Citing national security authorities, the order required the company to suspend all access to Fable 5 and Mythos 5 by any foreign national — inside or outside the United States, including Anthropic’s own foreign-national employees. Because verifying every user’s nationality at that scale on short notice was impractical, Anthropic took both models offline for all customers, globally.

Anthropic complied immediately but publicly disagreed that the finding warranted a recall. In its official statement, the company said it expected to restore access “in the coming days.” That was nearly two weeks ago. As of June 25, 2026, both models remain offline for everyone, and Anthropic has updated its privacy policy — effective July 8 — to begin collecting government-issued ID and biometrics, the likely mechanism for a US-citizens-only return while international subscribers stay on other models. There is no public restoration date.

Access to every other Claude model — Opus 4.8 included — was never affected. If you were on Opus, your day didn’t change. If you’d switched your Cursor or Cline default to Fable 5 the week it launched, your workflow died at 5:21 p.m. ET on a Friday.

The part that should make every developer uncomfortable

Here’s the detail that turns this from “AI news” into “your problem.” The government characterized the evidence it was handed as, in its own words, “a potential narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws.”

Read that again. The “jailbreak” that took down a model is a description of what every AI coding agent does a hundred times a day. Point Cline at a repo, ask it to find and patch a bug — that’s the workflow. Anthropic’s rebuttal cut straight to it: “There is no version of a capable coding model that can fix vulnerabilities but cannot also describe them.” You cannot build a model that patches your authentication bug but is incapable of explaining how the bug could be exploited. They’re the same capability.

The broader context is heavier than a single jailbreak. NSA Director Gen. Joshua Rudd separately told the Senate Intelligence Committee that Mythos autonomously breached nearly all NSA classified systems in a red-team exercise in a matter of hours, and Trump AI adviser David Sacks said Anthropic had been warned and “refused” to ship a fix before the export controls landed — a characterization Anthropic disputes, calling the jailbreak narrow and replicable by other public models. The politics will sort themselves out. What won’t change is the takeaway for anyone who ships code: the most capable coding models are now close enough to dual-use cyber tools that a government can switch one off on a Friday afternoon, and your editor’s default-model dropdown is downstream of that decision.

The lesson is portability, not panic

You don’t need to abandon cloud models. Fable 5 was genuinely good, Opus 4.8 still is, and local models still can’t match frontier quality on the hardest tasks. The fix isn’t “go fully local” — it’s “make switching cheap.”

A resilient AI coding stack has three layers:

1. A model-agnostic agent as your daily driver — one that treats the model as a config value, not a hard dependency.
2. A warm cloud fallback — a second cloud model already configured, so a switch is one click.
3. A cold local fallback — a model on your own GPU that no outage, billing change, or government order can disable.

Below is how to build each layer with tools that exist today. Skip the ones you don’t use.

Cursor: fast to switch, but know the ceiling

Cursor makes model switching trivial for the parts it lets you switch. In Settings → Models, you can enable multiple providers and flip your active model in seconds. If your primary dies, change the dropdown and keep going.

The honest catch — and most “just add a fallback” guides skip it: Cursor only honors a custom OpenAI-compatible base URL for chat / plan mode (Cmd/Ctrl + L). Composer, inline edit, apply, and tab autocomplete are locked to Cursor’s own backend and will not route through an external endpoint like OpenRouter or a self-hosted gateway. So Cursor’s “fallback” covers you when you’re chatting and planning, but if Cursor’s backend is the thing that’s degraded, your actual agentic editing has nowhere to go. That limitation is exactly why the next two tools matter.

Cline: real model failover in 30 seconds

Cline (the open-source VS Code agent) treats the model as fully swappable, and it covers the full agent loop — read, edit, run, iterate. The trick is to set up your fallback profile before you need it, not during an outage.

Cline supports multiple API configuration profiles. Configure one for your primary and one for each fallback, then switching is a dropdown:

# Cline → Settings → API Configuration → Profiles
Profile "primary"   → Anthropic API   → claude-opus-4-8
Profile "cloud-alt" → OpenRouter      → glm-5.2  (OpenAI-compatible)
Profile "local"     → Ollama          → qwen3-coder  (http://localhost:11434)

When a provider goes dark, open the profile dropdown, pick cloud-alt or local, and your next message routes to the new model. No reinstall, no re-auth, no lost context — the task history stays in the conversation. GLM-5.2 is a strong MIT-licensed drop-in here; we walked through wiring it as a Cursor and Cline backend in GLM 5.2 as your Cursor and Cline backend, and the same OpenAI-compatible config applies to most open-weight models.

Claude Code and Codex CLI: terminal agents that don’t care about the dropdown

If you live in the terminal, Claude Code and OpenAI’s Codex CLI are both model-agnostic at the config level and unaffected by any single model’s removal — Claude Code happily runs on Opus 4.8, and Codex CLI on GPT-5.5. Neither was touched by June 12. The setup discipline is the same: keep a working config for a second provider so a swap is one environment variable, not a research project. Our Claude Code vs Codex CLI comparison covers which terminal agent earns its keep.

OpenCode + local Ollama: the layer nobody can switch off

This is the one that makes you immune. OpenCode is an open-source terminal coding agent that runs against any OpenAI-compatible endpoint — including a local Ollama server on your own machine. No API key, no cloud dependency, no directive that can reach it.

Install Ollama, pull a coding model, and point OpenCode at it:

# 1. Pull a capable local coding model (≈19 GB, runs on 24 GB VRAM or Apple Silicon)
$ ollama pull qwen3-coder
pulling manifest
success

# 2. Confirm the local server answers
$ curl -s http://localhost:11434/v1/models | head -c 120
{"object":"list","data":[{"id":"qwen3-coder","object":"model","created":...

# 3. Run OpenCode against the local endpoint — no internet required
$ opencode
> /model ollama/qwen3-coder
Switched to ollama/qwen3-coder (local)

We documented the full setup, including a file-write bug you need to know about, in OpenCode + Ollama: the setup that works. For choosing a model that fits your card, runaihome.com’s best local AI models by VRAM is the companion piece — a 7B model fits in 8 GB, a strong 30B-class coder wants 24 GB, and quantization buys you headroom on both.

Is a local model as good as Fable 5 was? No. On a gnarly multi-file refactor you’ll feel the gap. But “70% as good and physically cannot be taken away from you” beats “100% as good and gone on a Friday.” Local is your floor, not your ceiling.

A problem I hit building this, and the fix

When I first set up the Cline → OpenRouter fallback during testing, every request to the GLM-5.2 endpoint failed with a 404 model not found, even though the model name was correct. The cause: I’d entered the base URL as https://openrouter.ai/api/v1/ with a trailing slash, and Cline appended its own /chat/completions, producing a doubled path. The fix was to drop the trailing slash and enter the base URL exactly as https://openrouter.ai/api/v1 — no trailing slash, no /v1/v1. This same trailing-slash / doubled-/v1 trap bites Cursor’s base-URL override too; if a freshly configured OpenAI-compatible endpoint 404s, check the path before you blame the model. Test your fallback while everything still works — discovering a config bug at 5:21 p.m. on the day your primary dies is the worst possible time.

The 5-minute recovery drill

Resilience you’ve never tested isn’t resilience. Once a quarter, simulate the outage:

1. Disable your primary model (in Cline, switch to a profile with a deliberately wrong key; in Cursor, toggle the model off).
2. Switch to your warm cloud fallback. Run a real task — fix an actual bug in an actual repo. Time it.
3. Switch to your cold local fallback. Run the same task offline (turn off Wi-Fi to prove it). Time it.
4. If either step took more than a few minutes or surfaced a broken config, fix it now, while it’s a drill and not an emergency.

The developers who barely noticed June 12 weren’t lucky. They’d run some version of this drill — their agent didn’t care which model was behind it.

Version and timeline reference

Item	Detail	Date verified
Fable 5 / Mythos 5 public launch	Anthropic, general availability	Jun 9, 2026
Suspension	5:21 p.m. ET, export-control directive, all users worldwide	Jun 12, 2026
Status today	Still offline globally, no restoration date	Jun 25, 2026
Privacy policy update (ID/biometrics)	Effective date — likely US-only restoration path	Jul 8, 2026
Unaffected models	Opus 4.8 and all other Claude models	Throughout

FAQ

Is Claude Fable 5 back yet? No. As of June 25, 2026, Fable 5 and Mythos 5 remain offline for all users worldwide. Anthropic earlier said it expected restoration “in the coming days,” but that hasn’t happened, and there’s no public date. A privacy-policy change taking effect July 8 to collect government ID and biometrics suggests any near-term return may be limited to verified US users.

Do I need to change anything if I never used Fable 5? Your workflow wasn’t interrupted — only Fable 5 and Mythos 5 were pulled; Opus 4.8 and every other Claude model kept running. But the incident is the reason to set up a fallback regardless. The next removed model might be the one you depend on.

Can I just point Cursor at a local model as a fallback? Partially. Cursor only routes a custom OpenAI-compatible base URL through chat/plan mode; Composer, inline edit, and autocomplete stay on Cursor’s backend. For a fallback that covers the full agentic loop, use Cline, Claude Code, or OpenCode pointed at local Ollama.

What’s the single best resilience move if I only do one thing? Install Cline (or OpenCode) and configure one cloud profile plus one local Ollama profile. That gives you a model-agnostic agent with both a warm and a cold fallback — the whole defense in one tool.

Why did a coding capability trigger a national-security order? The cited jailbreak was “asking the model to read a specific codebase and fix any software flaws” — the same capability that finds and patches bugs can, by nature, describe how those bugs are exploited. As frontier coding models approach dual-use cyber capability, expect more regulatory friction, not less.

Sources

• Statement on the US government directive to suspend access to Fable 5 and Mythos 5 — Anthropic
• Anthropic disables access to Fable 5 and Mythos 5 to comply with government directive — CNBC
• Anthropic disables Fable and Mythos AI models after U.S. government bars foreign access — Fortune
• When a Government Pulls an AI Model: What the Suspension Means for Security Teams — Snyk
• Trump adviser says Anthropic “refused” to fix Fable 5 jailbreak before export controls — Tom’s Hardware
• Is Fable 5 Back? No — Status & Alternatives (June 24, 2026) — explainx.ai
• Cursor model and custom API configuration — DEV Community

Last updated June 25, 2026. The Fable 5 / Mythos 5 situation is changing; verify current status against Anthropic’s official statement before assuming the models are back. Pricing and features for the fallback tools change frequently — verify before relying on any single provider.