Update, June 16, 2026: Fable 5 is suspended, so the fix in this post no longer works. On Friday June 12 at 5:21pm ET, Anthropic received a US government export-control directive ordering it to bar all access to Fable 5 and Mythos 5 by any foreign national, inside or outside the United States, including its own foreign-national employees. There is no way to filter foreign nationals out of US traffic in real time, so Anthropic disabled both models for everyone the same day. The practical effect: typing
/model claude-fable-5now returns a model-not-found or unavailable error no matter your plan, and you are effectively back on Opus 4.8. Every other Anthropic model stays online and unaffected. The government invoked national security authorities and did not publish specifics, though Anthropic understands the concern is a jailbreak of Fable 5, and reporting ties it to research that pulled cyberattack-relevant information past the model's safeguards. Anthropic says it is working to restore access. I am leaving the original launch-day post below as a record of how the switch worked, and it should apply again if access returns. Citations are in the fact-check notes at the end.
Anthropic released a new top model this morning, June 9, 2026. It is called Claude Fable 5, it sits in a new tier above Opus, and if you opened Claude Code and typed /model expecting to see it in the list, you got the same thing I did: Opus, Sonnet, Haiku, and no Fable anywhere.
I ran claude update. Already on the latest version (2.1.170 as I write this). Still no Fable in the picker.
This is not a bug and your install is not broken. Here is what's going on and the fix, which takes about five seconds.
What Fable 5 actually is
Quick orientation, because the naming is new. Anthropic announced two models together:
- Claude Fable 5 is the one you can use. It is a Mythos-class model, a capability tier above Opus, wrapped in safety classifiers that make it suitable for general release. Anthropic says it leads on nearly every benchmark they tested, with software engineering and scientific research called out specifically.
- Claude Mythos 5 is the same underlying model with the safeguards lifted. It is restricted to approved users, currently cybersecurity professionals through Anthropic's trusted access program, with biomedical researchers next.
The practical difference for you and me: Fable 5 is generally available on the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. When you send Fable 5 a query on one of the gated topics, the response comes from Claude Opus 4.8 instead. For normal business and coding work you will never notice. You are getting the stronger model with a fallback on a narrow set of subjects.
Pricing is $10 per million input tokens and $50 per million output. For reference, that is double Opus 4.8's standard rate of $5 and $25, and exactly equal to Opus 4.8's fast-mode rate. More on whether that trade is worth it below.
Why /model doesn't list it
The /model picker in Claude Code is not a live query against Anthropic's model catalog. It is a list baked into the CLI build you have installed. Your build shipped before Fable 5 existed, so the picker has never heard of it. claude update can't help on release day because there is no newer build to update to yet. The picker will catch up when the next CLI release ships.
The part most people miss: the picker is just a convenience menu. Claude Code will happily use any valid model ID you hand it directly.
The fix: type the ID
Inside any session, give /model the full model ID as an argument instead of picking from the menu:
/model claude-fable-5
That's it. Claude Code confirms with "Set model to claude-fable-5 and saved as your default for new sessions." You are now running Fable 5, today, on a CLI build that predates the model.
Three other routes to the same place, depending on how you like to work:
At launch, with a flag:
claude --model claude-fable-5
As an environment variable (PowerShell shown; add to your profile to persist):
$env:ANTHROPIC_MODEL = "claude-fable-5"
Pinned in settings.json (~/.claude/settings.json), which is the right answer if you want it durable across machines and sessions:
{ "model": "claude-fable-5" }
One prerequisite none of these can route around: your account has to have Fable 5 access provisioned. If you type the ID and get a model-not-found or permission error, that is an access problem on your plan or org, not a configuration problem on your machine. Nothing to fix locally; check your plan tier or ask your admin.
How to confirm it took
Type /model again with no argument. The status line shows the active model. You can also just ask the session what model it is, though the honest check is the status line, since a model's self-report can lag what the harness actually routed.
If you pinned it as default and later want back on Opus 4.8, same trick in reverse: /model opus or pick from the menu as usual.
Should you actually switch?
The under-$100 stack logic applies here the same way it applies everywhere else: pay for capability where capability is the bottleneck, and nowhere else.
At $10/$50, Fable 5 costs the same as running Opus 4.8 in fast mode and twice standard Opus. If your sessions are mostly routine, glue work, file edits, short questions, the honest answer is that Opus 4.8 was not your bottleneck and doubling your token rate buys you little. Keep your default where it is and reach for Fable when the task is genuinely hard: a thorny architecture decision, a research synthesis, a debugging session that has eaten an afternoon.
The pattern I would actually run: leave the cheap model as your session default, and switch with /model claude-fable-5 for the single hard task, then switch back. The five-second command makes per-task routing practical in a way it wasn't when changing models meant editing config files. If you have been following the routing approach from my model routing post, Fable 5 slots in as a new top rung on the ladder, not a replacement for the whole ladder.
And before you move any real workload onto it, run your own validation pass rather than trusting launch-day benchmarks. The checklist in validate a model before you upgrade was written for exactly this morning.
One more honest note: a model released today is past the knowledge cutoff of every assistant currently running, including the one helping you set it up. When I first asked about Fable 5 this morning, my own session confidently told me no such model existed, then checked the announcement and corrected itself. Verify against the source, not the model's memory. That habit costs nothing and catches a lot.
If you are building a business on top of tools like this rather than just using them, the playbook for shipping something small and paid on a stack this cheap is the whole subject of The $97 Launch.
What actually got it pulled, and why the story is messier than "jailbreak" sounds
The directive named a jailbreak, and Anthropic's statement said the government shared no specifics. The reporting since is worth reading closely, because the bypass is not the exotic exploit the word implies.
By the accounts published so far, researchers pointed Fable 5 at a specific codebase and asked it to read the code and fix its flaws, then used a series of prompts to steer it toward information framed as useful for cyberattacks. Anthropic characterized what came out as a small number of previously known, minor vulnerabilities that all appear relatively simple. Put plainly, the trigger looks close to a routine defensive-security task: give a strong model some code and ask it to find and fix problems.
Two things about how that reached a federal order deserve to be said directly. The push came from Amazon, a direct competitor, whose CEO raised it with the White House, and David Sacks described the source as a "highly credible trusted partner of both Anthropic and the US government." Separately, IBM X-Force researchers had been arguing before any of this that Fable 5's guardrails were too aggressive for legitimate defensive work. So the same model was pulled for showing the kind of capability blue teams lean on daily, not long after complaints that its safeguards were already too tight. Hold both readings at once: this was either a real capability leak or an overreaction amplified by competitive and political pressure, and the public evidence does not settle it. Treat "jailbreak" as a claim to examine, not a fact to accept.
Why a narrow, contested bypass was enough to pull the model worldwide comes back to what Fable 5 is. It is Mythos-class capability wrapped in safety classifiers, and those classifiers are the whole reason it can ship to the public. A method that gets past them, even a limited one, reads as frontier capability without the guardrails, and that is what export-control authorities respond to regardless of how serious any single finding looks.
The practical lesson is less about this model and more about the shape of the risk. Model availability is now a variable you do not control. Something you depend on can disappear on a Friday afternoon because of a regulator, a rival's research, or a finding whose details you never see. The security teams writing up the suspension converge on the same advice: do not let any single hosted model become a hard dependency, so build redundancy and graceful fallbacks into anything that matters; keep an inventory of where AI components actually sit in your stack so you know the blast radius when one vanishes; scan AI-generated code by default rather than as an afterthought; and prefer guardrails and monitoring over kill switches, constraining actions and watching behavior instead of cutting the whole thing off. The five-second model switch this post is built around is a small version of that same principle: keep routing flexible so swapping the top rung is one line, not a rebuild.
Fact-check notes and sources
- Claude Fable 5 and Claude Mythos 5 announcement, June 9, 2026: anthropic.com/news/claude-fable-5-mythos-5. Model tier, safeguard design, Opus 4.8 fallback behavior, and availability claims are from this page.
- Pricing of $10 per million input tokens and $50 per million output tokens for both models: same announcement page.
- API model ID
claude-fable-5: stated on the announcement page ("use claude-fable-5 via the Claude API") and confirmed working in Claude Code 2.1.170 by the author on June 9, 2026. - General availability on the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry; Mythos 5 limited availability through Project Glasswing: Anthropic announcement and same-day coverage by TechCrunch and NBC News.
- Opus 4.8 standard pricing of $5/$25 and fast-mode pricing of $10/$50 per million tokens: Anthropic Opus 4.8 release materials, May 2026.
- The
/model <full-id>behavior and "saved as your default for new sessions" confirmation: observed directly in Claude Code 2.1.170 on Windows, June 9, 2026. - Suspension of Fable 5 and Mythos 5 (added June 16, 2026): primary source is Anthropic's own statement, anthropic.com/news/fable-mythos-access. It records the US government export-control directive received June 12, 2026 at 5:21pm ET, the bar on "all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States," the decision to "abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance," and that "access to all other Anthropic models will not be affected."
- Date, worldwide scope, and the jailbreak / national-security framing corroborated by CNBC, Al Jazeera, Tom's Hardware, and The Hacker News, June 12 to 13, 2026.
- The jailbreak analysis (what the bypass actually was, Anthropic's "minor, previously known" characterization, the Amazon-led research and White House escalation, and the security-team takeaways): The Hacker News for the "read a specific codebase and fix any software flaws" description, the "small number of previously known, minor vulnerabilities" quote, and the Amazon / Andy Jassy / David Sacks attribution; Snyk for the IBM X-Force "too aggressive guardrails" point and the model-redundancy, scan-by-default, and guardrails-over-kill-switches recommendations. Published June 13 to 14, 2026.
