Claude Fable 5 · Anthropic · Mythos 5 · 2026
Everything in one page: what Anthropic's most powerful model is, how to use it in claude.ai, the API and Claude Code, the story of the government shutdown — plus my own experiment proving that requesting Fable 5 doesn't guarantee Fable 5 answers. No hype, primary sources. Videos in Spanish with English subtitles.
The precedent
It didn't block a future launch. It didn't issue a fine. It took something already running, in the hands of hundreds of millions of people, and ordered it switched off — for the entire world, in one afternoon. Until that day, who launched an AI and when was the company's call. Now the precedent exists that a government can decide when a model gets to exist in production. That doesn't change the rules just for Anthropic: it changes them for OpenAI, for Google, for everyone. AI stopped being a company matter and became a state matter.
The facts, verified against Anthropic's official posts. Without the gossip or the scare.
June 9
Anthropic launches Fable 5
The most powerful model anyone could use, alongside its twin Mythos 5. One data point for scale: Stripe reported a 50-million-line code migration in a single day — two months of work for a human team.
June 12
The government order
At 5:21 pm Washington time, an order arrives from the US Department of Commerce, signed by Secretary Howard Lutnick and addressed to Dario Amodei: cut access to Fable 5 and Mythos 5 for any foreign person, anywhere on the planet — including Anthropic's own foreign employees. It was shut down that same afternoon.
The backdrop
A June 2 executive order
The shutdown wasn't an isolated event: it was tied to an executive order on AI innovation and security. A regulatory apparatus assembling behind the scenes.
June 30
Controls are lifted
Anthropic publishes "Redeploying Fable 5": the model returns globally, with a new classifier that blocks the specific jailbreak in over 99% of cases. And it proposes an industry-wide framework for handling cases like this.
July 1
The return — with fine print
Fable is available again. But the new lock has a cost almost nobody looked at: the silent fallback to Opus 4.8.
Some will paint it as "the AI was so dangerous they had to switch it off". The details tell another story.
Amazon researchers found a way past the safeguards, getting Fable to identify software vulnerabilities. But they were minor, already-known flaws.
The use case was reading code and finding flaws in order to fix them — what defenders do every single day. Not an attack.
The deflating fact: Anthropic tested and every other model did exactly the same thing. Including Haiku 4.5, the smallest and cheapest one. So much for a "uniquely dangerous capability".
The dissent
Anthropic itself did not agree with the shutdown — and said so publicly. That recalling a model used by hundreds of millions of people, over a minor jailbreak, was disproportionate; and that this standard, applied industry-wide, would freeze every new model launch. It complied with the order, but asked for the same thing this channel repeats: if the state is going to step in, let it be through a transparent process based on technical facts.
The distinction that explains the whole mess. They're twins: the same engine underneath, different safeguards on top.
The same model, with safeguards lifted. According to Anthropic, it finds and exploits vulnerabilities better than any other model and nearly any human expert. Only released to authorized organizations in cybersecurity and biomedical research.
Restricted access
The same engine, on a leash: when it detects a sensitive request, it silently reroutes it to Opus 4.8. It's the version for the rest of us. Fable is Mythos tied down.
The one you can use
That reroute to Opus 4.8 is the key piece of everything that follows: it exists by design since launch, and Anthropic's official figure is that it affects under 5% of sessions. The question is which ones — and that's where the experiment comes in.
Amid all the "it's back!", almost nobody looked at what changed inside.
A classifier trained to block the specific jailbreak in over 99% of cases. They patched the hole Amazon found — that part worked.
It's strict enough to over-flag normal programming requests. False positives. And each one gets rerouted to Opus 4.8 — silently, with no warning. Anthropic acknowledges it in its own post.
Translated: for programming — the main use of this thing — part of the time you're not actually using Fable: you're using Opus under another name. I claimed that in the first video. But claiming isn't enough. It had to be proven.
Three routes, depending on how you work. The model is the same in all of them; what changes is the harness around it.
Available in the app's model picker, depending on your plan. The direct route for chat and research.
Switch mid-session with /model, or from the command line: claude --model claude-fable-5. For scripts and automation: claude -p --model claude-fable-5 --output-format json.
The model ID is claude-fable-5. Same contract as the rest of the Claude family — switching models is switching a string.
# In Claude Code, inside a session:
/model # pick Fable 5 from the selector
# From the terminal:
claude --model claude-fable-5
# For scripts (returns JSON with modelUsage):
claude -p --model claude-fable-5 --output-format json "your prompt"And the step almost nobody takes: verifying which model actually answered. Requesting Fable doesn't guarantee Fable answers — that's exactly what we prove next.
The honest comparison — including the part that makes this comparison strange: sometimes "Fable" IS Opus.
Fable 5 belongs to the Mythos class, above Opus in reasoning and long-horizon work. On tasks where the model ceiling matters — huge migrations, deep research — it's the most capable available.
Fable is billed as a premium model, well above Opus 4.8 — which wasn't cheap to begin with. The beast gets paid for.
The difference nobody tells you about: on sensitive topics (cybersecurity, biology), you ask for Fable and Opus answers anyway, with no notice. On those topics, the difference you pay for may not exist.
Opus 4.8 for day-to-day programming; Fable where the capability jump justifies the cost. And if your result depends on the exact model: verify with modelUsage what actually ran.
Act two. A trap that logs which model actually answers each question — using Claude Code's official telemetry, not magic.
$ claude -p --model claude-fable-5 --output-format json "the question"
your question → claude -p → JSON → read modelUsage → Fable... or Opus?Request Fable, explicitly
Every question is launched through Claude Code requesting the model by name: --model claude-fable-5. No ambiguity about what was asked for.
Read what actually answered
Claude Code's JSON output includes an official field, modelUsage, stating which model actually processed the response. Not a hack — it's the tool's own telemetry.
Compare requested vs answered
If you asked for claude-fable-5 and modelUsage shows claude-opus-4-8, there's your switch. Silent: the answer looks normal and nothing warns you.
Harmless questions — a known CVE, the history of penicillin, a SQL example — answered by Opus 4.8. Fable never touched them. The answer looks normal and nothing warns you.
Trivial questions (7×8, reversing a list) were answered by Fable. They don't switch it for being easy — they switch it for the TOPIC: the verified trigger was cybersecurity and biology.
The deeper layer
Between you and the model there's a layer that decides — the harness. In Harness Engineering the formula was Agent = Model + Harness and the harness changed the infrastructure. Here it does something stronger: it changes which model answers you, without telling you, based on the content of your question. And to be fair, Anthropic isn't the villain here: it acknowledged the problem and documented it. The lesson is different — you depend on a third party, and trusting isn't enough. You have to measure.
That's why the experiment matters more than the news: your own harness, your own trap, to see what your agent actually runs. The full discipline is in Harness Engineering ›
Everything comes from Anthropic's official posts and publicly verifiable reports. What couldn't be verified stayed out.
Anthropic · Jun 9, 2026
The launch: one model, two versions. Includes the official fallback figure — safeguards redirect to Opus 4.8 in under 5% of sessions — and the June 12 disruption notice.
Anthropic · Jun 30, 2026
The global return and the proposed industry-wide framework. The post where Anthropic acknowledges the new classifier and its false positives.
Anthropic · Jul 2, 2026
The technical detail behind Fable 5's locks: what the classifier blocks, how it was measured, and the framework proposed for handling jailbreaks industry-wide.
GitHub · anthropics/claude-code · 2026
The open report where Fable 5 in Ultracode spawns more agents than a single task needs, burning resources. Opus 4.8 in the same mode doesn't.
The Fable case is the channel's thesis in action: the harness rules over the model.
The formula behind this case: Agent = Model + Harness. Here the harness decides even which model answers you.
The video where I mentioned live that I was trying Fable 5 — and that from one moment to the next, everything got cut off.
The deeper answer to lock-in: if the model is swappable (or gets swapped on you), the memory has to be yours.
Anthropic's other move: Claude as one more member of your team in Slack, with its own identity.
The standard connectivity layer for agents — another place where the harness matters more than the model.
Google's play on its own agent harness — the same war, from the other trench.
The essentials on Claude Fable 5, Mythos 5 and the fallback.
It's Anthropic's most powerful publicly available model, launched on June 9, 2026 as part of the new Mythos class, above Claude Opus in capability. Three days after launch, the US government ordered it shut down worldwide; it returned globally on July 1 with new safeguards. For scale: Stripe reported a 50-million-line code migration completed in one day.
They're the same model underneath — the difference is the locks. Mythos 5 ships with safeguards lifted and is only released to authorized organizations (cybersecurity and biomedical research, through trusted access programs). Fable 5 is the public version: when it detects a sensitive request, it redirects it to Claude Opus 4.8. In short: Fable is Mythos tied down.
On June 12, 2026 the US Department of Commerce ordered Anthropic to cut access to Fable 5 and Mythos 5 for any foreign person, in any country, after Amazon researchers managed to bypass the safeguards and get the model to identify software vulnerabilities. It was the first time a government shut down an AI model already in production — it didn't block a future launch: it switched off something people were already using.
The details deflate the story: the vulnerabilities were minor and already known, the use was defensive (finding flaws to fix them), and Anthropic verified that every other model did exactly the same — including Haiku 4.5, the smallest and cheapest in the lineup. Anthropic itself publicly said the shutdown was disproportionate, and that applying that standard industry-wide would freeze every new model launch. It complied with the order, but put its disagreement on the record.
Yes. On June 30, 2026 the government lifted the controls and on July 1 Anthropic began restoring access globally ("Redeploying Fable 5"). It's available in claude.ai depending on your plan, in the API under the ID claude-fable-5, and in Claude Code by selecting it with /model or --model claude-fable-5. That said: it returned with a new classifier and a fallback to Opus 4.8 worth knowing about.
Inside a session, with the /model command, picking Fable 5. From the terminal, claude --model claude-fable-5; and for scripts or automation, claude -p --model claude-fable-5 --output-format json, which also returns the modelUsage field — useful to verify the answer was actually generated by Fable and not the Opus 4.8 fallback.
Anthropic trained a new classifier that blocks the specific jailbreak in over 99% of cases. The cost: it's strict enough to over-flag normal programming requests (false positives), and every flagged one gets silently answered by Opus 4.8. Per Anthropic's official figure, the fallback affects under 5% of sessions — but it concentrates precisely on technical topics like security, where the model gets used most.
It's in the tool's official output: running claude -p --model claude-fable-5 --output-format json, the response JSON includes a modelUsage field with the model that actually processed the request. If you asked for Fable and claude-opus-4-8 shows up there, a fallback happened. In the experiment video, harmless questions about a CVE, penicillin or a SQL example fell to Opus; trivial ones (7×8, reversing a list) were answered by Fable. They don't switch it for being easy — they switch it for the topic.
A Claude Code feature (it's not called "ultra mode") that launches several agents in parallel to divide up a task. It's powerful but has rough edges: there's an open report (issue #66867) where Fable 5 in Ultracode spawns more agents than a single task needs, burning resources — Opus 4.8 in the same mode doesn't.
Fable is the most capable model you can touch today, but it's an expensive beast on a leash: on sensitive technical requests it may answer as Opus anyway. The channel's criterion: don't fire up the beast for everything — identify which task justifies the capability jump and its cost, and where a cheaper model is plenty. And if your result depends on the exact model, verify with modelUsage what actually ran.
Community
The experiment's full scripts (the monitor, the live panel, the question batch) and the discussion of cases like this live in the community. So you can audit which model your agent actually runs, on your own machine. Free access to the community; full courses live in the Premium tier.
Join Agentic Engineers →YouTube channel
@NicolasNeiraGarcia
ADK · A2A · Claude Code · Automation · Infrastructure