Claude Fable 5 and Mythos 5: Launched, Then Pulled

Last updated: June 2026

What are Claude Fable 5 and Claude Mythos 5?

They are the same underlying model, shipped in two versions. Claude Fable 5 is the general-availability version, with safeguards in place; Claude Mythos 5 is the same model with some safeguards lifted for vetted partners. Anthropic released both on June 9, 2026 and described the tier as “Mythos-class” — its first models positioned above the Opus class in raw capability.

This is the model family the community had been anticipating since the pre-launch roadmap chatter. Our earlier explainer, what Anthropic ships after Opus 4.8, was written before the launch; Fable 5 and Mythos 5 are the confirmed reality of that “Mythos” tier. They also slot in just above Claude Opus 4.8, which is the model both versions silently fall back to whenever a request hits a restricted topic.

What does “Mythos-class” actually mean?

“Mythos-class” is Anthropic’s name for a capability tier sitting above the Opus class. In Anthropic’s own framing, Fable 5 is “state-of-the-art on nearly all tested benchmarks,” with the performance gap over previous models widening on longer, more complex, multi-step work. The headline difference is autonomy duration: Mythos-class models can run coherently for far longer before drifting — the trait that matters most for agentic coding and research.

The naming split is the practical part. Fable = the safe, general-use face of the model. Mythos = the same weights with specific safeguards removed for trusted partners (for example, cyber safeguards lifted for “Project Glasswing” partners, and biology/chemistry safeguards lifted for select researchers). If you used the public model, you used Fable 5.

Capability and pricing figures are Anthropic-reported and were valid at the June 9 launch; treat benchmark claims as vendor-reported unless independently audited.

How powerful is Claude Fable 5?

By Anthropic’s published results, it set a new bar across coding, knowledge work, vision and long-context tasks. These are vendor-reported metrics, so read them as directional rather than audited:

• Software engineering: highest score on Cognition’s FrontierCode evaluation among frontier models; Stripe reported a 50-million-line Ruby codebase migration done in a day rather than the roughly two months a manual effort would take. See our running ranking of the best AI coding assistants in 2026 for where this fits.
• Knowledge work: top score on Hebbia’s Finance Benchmark for senior-level reasoning over documents, charts and tables.
• Vision: rebuilt web-app source code from screenshots and finished Pokémon FireRed using vision only with a minimal harness.
• Long-context memory: on the game Slay the Spire, persistent file-based memory roughly tripled performance versus Opus 4.8.

What did Mythos 5 do in scientific research?

This is where the Mythos version — the one with research safeguards lifted — produced its most striking, and least independently verified, results. Every figure below is Anthropic’s own:

• Genomics: over roughly a week of largely autonomous work, Mythos 5 assembled single-cell data for millions of cells across 138 animal species and built a custom machine-learning model that, Anthropic says, outperformed a recent Science-journal result despite being about 100× smaller.
• Protein and drug design: internal experts reported a roughly 10× speed-up in a drug-design workflow; on 14 protein targets, 9 yielded strong candidate molecules worth investigating, and on selected tasks the model “matched or beat skilled human operators.” Its AAV (adeno-associated virus) designs reportedly outperformed dedicated protein language models.
• Novel hypotheses: Anthropic calls it the first model to “consistently produce novel, compelling scientific hypotheses.” In blind comparisons, scientists preferred its hypotheses to Opus-class output about 80% of the time, and one hypothesis about an E. coli protein was later corroborated by an independent lab study.

If even half of this holds up, it is the clearest sign yet that frontier models are crossing from “useful research assistant” into “produces results worth publishing.” That is precisely why the biology and chemistry safeguards on the public Fable 5 mattered — and precisely the capability the export order is trying to contain.

What did the API and pricing look like?

At launch, Fable 5 was a normal API model — same shape as any other Claude call, just a new model ID and a higher price ($10 per million input tokens, $50 per million output tokens, which Anthropic noted was less than half the cost of the earlier Mythos Preview). A minimal call looked like this:

from anthropic import Anthropic

client = Anthropic()

# Fable 5 — the public, fully-safeguarded Mythos-class model.
# As of mid-June 2026 this call returns an error: access is suspended.
resp = client.messages.create(
    model="claude-fable-5",
    max_tokens=2048,
    messages=[{"role": "user", "content": "Migrate this module to async."}],
)
print(resp.content[0].text)
# Pricing at launch: $10 / 1M input tokens, $50 / 1M output tokens.

What are the limitations and safeguards?

Even before the suspension, Fable 5 shipped with hard guardrails — the “limitations” worth knowing are mostly deliberate safety design, not capability gaps:

• Opus 4.8 fallback: on restricted topics — chiefly cybersecurity and biology/chemistry — Fable 5 silently hands the query to Claude Opus 4.8 instead. Anthropic says this triggered in under 5% of sessions, and users are told when it happens.
• Anti-distillation: a classifier blocks attempts to extract the model’s capabilities to train competing models.
• Self-throttling on frontier-AI work (reported): observers reported that Fable 5 appeared to quietly hold back when it detected a user working on frontier AI development — unlike the topic fallbacks, with no visible notice. Treat this as community observation rather than a documented feature.
• 30-day retention: all Mythos-class traffic was retained for 30 days for safety monitoring, with human access logged and deleted afterward in nearly all cases — not used for training. That is stricter logging than Anthropic’s standard tiers.
• Bio/chem gating: most biology and chemistry requests fell back to Opus 4.8; the unrestricted Mythos 5 behaviour was confined to a small set of vetted researchers.

Anthropic also rated Mythos 5’s misaligned-behaviour level as “low, and similar to Opus 4.8,” and said external red-teamers found no universal jailbreaks across 1,000+ hours of testing. That last point matters for the next section.

Why did the US government suspend Claude Fable 5 and Mythos 5?

Because of an export-control directive, not a safety failure Anthropic agreed with. On June 12, 2026 at 5:21pm ET, Anthropic received a US government order to suspend all access to Fable 5 and Mythos 5 “by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.” Because the company cannot reliably separate foreign nationals from everyone else in real time, the only way to comply was a hard global shutoff — so both models went dark for every user.

According to CNBC and other outlets, the directive came via a letter from Commerce Secretary Howard Lutnick, drafted with the Commerce Department’s Bureau of Industry and Security (BIS). The government cited national-security authorities but did not publicly detail the concern. Anthropic says officials pointed to a specific technique that could circumvent some of Fable 5’s narrow safeguards — and Anthropic publicly disagreed that a narrow potential jailbreak justifies recalling a commercial model already deployed to hundreds of millions of people.

How did we get here? Anthropic vs. the US government

The suspension did not come out of nowhere — it is the latest flashpoint in months of friction between Anthropic and the Trump administration:

• March 2026: the administration designated Anthropic a “supply chain risk.” Anthropic filed two lawsuits challenging the designation as unlawful retaliation for refusing to drop its restrictions on military and surveillance use of its models.
• April 2026: in a separate matter, a federal judge blocked a ban on US-government use of Anthropic’s technology, and the Department of Justice filed notice to appeal. This is distinct from the June export order, but it shows the same adversarial pattern.
• June 9–12, 2026: Fable 5 and Mythos 5 launch, then the export-control directive lands three days later and both go dark.

So the order reads less like a one-off safety call and more like the next move in an ongoing standoff. That same regulatory pressure shapes Anthropic’s broader trajectory, including the questions raised in our breakdown of the Anthropic S-1 IPO filing — for a frontier lab, government risk is no longer a footnote.

What does the suspension mean if you’re outside the US?

Right now, the same as for everyone else: no access. The directive specifically targets “any foreign national,” and because Anthropic cannot filter nationality in real time, it shut both models off globally. But the framing matters if you build outside the United States — this is an export-control measure tied to who you are, not a product recall over a shared bug.

If you were trialling Fable 5 from Europe or elsewhere, the practical moves are simple:

• Fail back to Opus 4.8. It is unaffected and was already Fable 5’s own fallback for restricted topics. For most production work the capability gap is real but not blocking — see what changed in Opus 4.8.
• Don’t hard-code claude-fable-5. Keep the model ID behind config so a restoration — or a permanent retirement — is a one-line change, not a redeploy.
• Treat regulatory exposure as a selection criterion. A model that can vanish on a Friday afternoon is a different procurement risk than one that can’t, which is part of why availability now sits alongside benchmarks in our ChatGPT vs Claude vs Gemini comparison.

What does this mean in practice?

Three takeaways that competitors covering only the suspension are missing:

1. The frontier briefly moved, then got locked in a vault. If Anthropic’s numbers hold, Fable 5 was a genuine step over Opus-class on long-horizon work. But it is now the first frontier model effectively quarantined by export policy — capability that exists, was sold, and then became inaccessible mid-contract. For teams that built pipelines around claude-fable-5 in its three live days, the lesson is brutal: a frontier model ID is now a geopolitical dependency, not just a technical one.

2. The Fable/Mythos split was the right design — and it didn’t save them. Anthropic separated the safe public model (Fable) from the unlocked partner model (Mythos) precisely to contain risk. The government still ordered both offline over a “narrow” Fable jailbreak. That tells you regulators are now reasoning about worst-case capability, not the safeguarded default — a shift every lab shipping agentic models should plan for.

3. “Most capable model” is no longer the only axis that matters. For a working developer in mid-2026, an always-available Opus 4.8 or a strong rival beats a suspended state-of-the-art model every time. If you are choosing a stack today, weight availability and regulatory exposure alongside benchmarks, not just leaderboard position.

Sources prioritise Anthropic’s primary disclosures and major business newsrooms (CNBC, NBC News, Bloomberg, Fortune). Benchmark and science figures are vendor-reported by Anthropic and not independently audited. The legal and access status was developing as of mid-June 2026; verify the current position before relying on it. Links accessed June 2026.