Claude Fable 5: Anthropic ships its Mythos-class frontier model — deliberately handicapped

Anthropic shipped two frontier models today, and by the company's own admission they are the same model. Claude Mythos 5 is the raw artifact — the most capable thing Anthropic has ever trained. Claude Fable 5 is what the rest of us get: identical weights, wrapped in safety classifiers that step in when a conversation drifts into territory Anthropic considers too dangerous for general release.

It is the first major launch where the headline model is openly advertised as the restricted variant, and Anthropic isn't being coy about it. The names even encode the relationship: fable comes from the Latin fabula, "that which is told" — the company's nod to the Greek mythos. Same story, different teller. The safeguards, Anthropic says, "are what distinguish the two models and are why we've given them different names."

What the governor actually does

Fable 5 carries three classifier systems:

• Cybersecurity — blocks exploitation assistance and offensive cyber operations. Anthropic says external red-teaming produced zero successful jailbreaks across more than 1,000 hours of testing.
• Biology and chemistry — blocks dual-use research with pandemic potential, including viral design. Anthropic concedes these deliberately err on the side of caution.
• Distillation — resists capability-extraction attempts aimed at cloning the model.

The interesting design choice is what happens when a classifier fires. Instead of a refusal, the request is silently re-routed to Claude Opus 4.8 — the previous flagship — which answers in Fable's place. Anthropic says this happens in under 5% of sessions on average. So Fable 5 doesn't say no; it just gets measurably dumber for a turn, and unless you're reading benchmark footnotes you might never notice.

The numbers — and the asterisks

On the open benchmarks, the Mythos class posts the strongest scores Anthropic has ever published:

Those asterisks are the story. Anthropic's methodology note says the published number is the higher of the two models, and that on starred benchmarks "Claude Fable 5 performs closer to Claude Opus 4.8 due to fallbacks." Translation: on cybersecurity- and biology-adjacent work, the model you can actually buy doesn't score 78% on ExploitBench — something nearer Opus 4.8's 40% does, because the classifier keeps swapping the frontier model out from under you. The capability exists in the weights you're calling. You're just not allowed to reach it.

Where the handicap bites — and where it doesn't

For most developers, the honest answer is: it mostly doesn't. The agentic-coding results that matter day to day — SWE-Bench Pro, FrontierCode, computer use, tool use, the knowledge-work evals — are unstarred. That 80.3% SWE-Bench Pro score, eleven points clear of Opus 4.8 and twenty-one clear of GPT 5.5, is the real Fable 5 you get over the API. Early access stories back it up: Stripe reports compressing months of engineering work into days.

The people who genuinely get the handicapped model are security researchers and red teams — exactly the crowd staring at that ExploitBench delta, where nearly half the score is being deliberately withheld — plus anyone working in wet-lab-adjacent biology. For them, Anthropic's answer is the unrestricted Mythos 5, gated behind Project Glasswing partnerships and invitations to vetted biology researchers, with a 30-day data-retention policy for safety monitoring. Behind that gate the results are striking: protein-design teams report 10x faster drug-design workflows, scientists preferred Mythos-generated molecular biology hypotheses about 80% of the time over Opus-class output, and an autonomous genomics run outperformed a published Science paper.

Pricing and availability

Fable 5 is live on the Claude API today at $10 per million input tokens and $50 per million output, with a staged rollout to subscription plans running June 9–23. Mythos 5 is application-only.

Whether you read this launch as responsible deployment or as paying frontier prices for a model with a governor bolted on probably depends on which side of the classifier you work. But credit where due: Anthropic published the asterisks, named the fallback, and told you exactly which model answers when the safety system decides you've asked the wrong question. A handicap you can see is a better deal than one you can't.