I Switched to Fable 5 on Day One. Four Days Later the Government Shut It Off.

Four days ago I published a day-one post on Claude Fable 5, Anthropic's first model tier above Opus, and I ended it with a warning I thought was about money: don't build a dependency on a model you might lose access to in two weeks. I was worried about the June 22 subscription cliff. I was worried about the wrong thing. On Friday afternoon — 5:21pm ET, June 12 — the US government issued an export-control directive ordering Anthropic to suspend all access to Fable 5 and its unrestricted sibling, Mythos 5. Anthropic is complying by disabling both models for every customer, globally. The model that's been running my Claude Code sessions since Monday is, as of this weekend, gone.

This is breaking, and it's thin. The directive landed late Friday; Anthropic said it would share more within 24 hours, and that window closes this evening as I write. So read everything below as a Saturday-morning snapshot — I'll mark what's confirmed versus what's still developing, and update when the rest lands. The part that matters for builders, though, is already settled: the model you were calling yesterday returns an error today.

What the directive actually says

Here's the confirmed core, in Anthropic's own framing. The US government, citing national security authorities, issued an export-control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national. To comply, Anthropic is disabling both models for all customers worldwide — there's no nationality switch at the API layer, so in practice everyone loses them. Every other Anthropic model — Opus 4.8, Sonnet 4.6, Haiku — is unaffected.

The trigger is the part that should make every builder of AI coding agents sit up straight. The government pointed to a capability it framed as a jailbreak risk: asking the model to read a codebase and identify software flaws. That's the whole thing. Anthropic calls it a narrow potential jailbreak and points out the same capability is widely available from other models (including OpenAI's GPT-5.5).

Sit with that. Reading a codebase and finding its flaws isn't an exotic exploit — it's the headline feature. It's the 50-million-line Stripe migration Anthropic put in its own launch material. It's my PR-reviewer agent. It's what a lot of us point these models at all day. And it's security-flavored, which puts it squarely in the zone where Fable 5 was designed to silently fall back to Opus 4.8 instead of answering. The most jailbreak-resistant model Anthropic says it's ever shipped — the one I noted held up across 400 turns of red-teaming — got pulled over the exact capability its safeguards were built to be careful about.

Anthropic doesn't agree with the call, and says so plainly. Its case: perfect jailbreak resistance is not currently possible for any model provider, the safeguards followed thousands of hours of red-teaming with government and third-party teams, and non-universal jailbreaks will always eventually surface. Apply this standard evenly, the company warns, and you'd essentially halt all new model deployments for all frontier model providers. That's the real fight, and it's bigger than one jailbreak: is 'a frontier model can find bugs in code' now an export-controlled capability? If the answer is yes, every lab is exposed.

One more thing worth naming, carefully. In the day-one post I noted that Mythos 5 — the unrestricted version — is the model Anthropic said it deployed with the US government, through a program it called Project Glasswing. So the same government that co-deployed Mythos is now invoking national-security export controls to switch it off. I don't know what to make of that yet, and I won't pretend otherwise. But it's on the record.

What breaks if you adopted it

If you took my day-one advice literally and switched, here's the blast radius. None of it is hypothetical — it's the checklist I worked through on my own setup this morning.

Claude Code defaults. If Fable 5 was your selected model, your sessions can't reach it anymore. Re-select Opus 4.8 with /model. The agent that helped research my day-one post was running on Fable; it's back on Opus now.

Hardcoded model IDs. Any API call with claude-fable-5 now fails — that's a disabled model, not a silent reroute. Anything in production carrying the string claude-fable-5 needs to point at claude-opus-4-8 today. And if you reached Fable through GitHub Copilot or Amazon Bedrock, where it had just gone GA, it's the same story — the suspension is global, not Anthropic-API-only.

In-flight evals and A/B tests. I had my LangChain + Mastra PR-reviewer running on Fable and Opus side by side for two weeks, chasing honest cost-per-task numbers. That experiment just lost an arm. If you were benchmarking against Fable, your baseline evaporated — archive what you've got, because the Fable runs aren't reproducible anymore.

The prompt cache, again. Switching models invalidates your prompt cache, so the first wave of requests after you revert to Opus rebuilds it from scratch — a one-time cost bump, the mirror image of the one you paid switching to Fable.

The clean path back to Opus 4.8

The migration back is mostly the day-one migration run in reverse, and there's a tidy symmetry in the gotcha. Remember the parameter that 400'd your requests on Fable — an explicit thinking: { type: "disabled" }? Opus 4.8 accepts it again. The thing that broke on the way in is legal on the way out.

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

const response = await client.messages.create({
  // was: model: "claude-fable-5"  <- now errors; the model is disabled
  model: "claude-opus-4-8",
  max_tokens: 16000,
  thinking: { type: "disabled" }, // legal again on Opus 4.8; 400s on Fable 5
  output_config: { effort: "high" },
  messages: [{ role: "user", content: "..." }],
});

Two more things to watch on the way back. First, Opus 4.8's minimum cacheable prefix is 4,096 tokens, up from Fable's 2,048 — so prompts in that 2K–4K band that quietly started caching on Fable will quietly stop caching on Opus. Watch your cache-hit rate, not just your error logs. Second, everything else you tuned — effort levels, task budgets, structured outputs, the 1M context window — is identical across the two, because under the hood Opus 4.8 was always Fable's own safety fallback. You're reverting to the model Fable was quietly handing your security queries to anyway.

In Claude Code specifically: /model, pick Opus 4.8, done. If you'd saved Fable as your default, reset it so new sessions don't open by reaching for a model that isn't there.

The honest part — building on a model that can vanish

I don't want to take a victory lap, because the lesson is more uncomfortable than 'I told you so.' I did warn, four days ago, to keep your Opus 4.8 path warm because you might lose Fable. I was right about the failure mode and completely wrong about the cause. I'd priced in a capacity-and-subscription cliff on June 22. I had not priced in an export-control order on a Friday. What made me keep a fallback was a guess about Anthropic's billing; what actually pulled the model was the US government. Being right for the wrong reason isn't a strategy.

So here's the real takeaway, and it's bigger than Fable: model access is now a regulatory variable, not just a pricing-and-capacity one. A frontier model can disappear overnight by government directive — no notice, no migration window — and the specific trigger here was a capability (read this code, find the flaws) that is the entire daily premise of AI coding agents. If that's genuinely export-controlled now, this won't be the last suspension, and it won't stay limited to Anthropic. The company's own warning — that this standard would halt all new model deployments across frontier providers — is either special pleading or a forecast. I genuinely don't know which yet.

What I'm actually doing this weekend: Claude Code is back on Opus 4.8. The PR-reviewer's Fable arm is archived, not deleted. And the thing that selects a model is, from today, a single environment variable with claude-opus-4-8 as the default and a documented swap path — not a string copy-pasted across a dozen call sites — because the next time a model evaporates, I want the fix to be one line, not an afternoon. The model-agnostic plumbing I used to file under over-engineering just became table stakes.

I'll update this post when Anthropic's promised details land: what 'access' precisely covers, whether there's any reinstatement path, and whether Mythos 5's government ties reshape any of it. Until then — revert, document your fallback, and assume the model under your agent is borrowed, not owned. Four days turned out to be a short loan.