I was on a call where a security engineer read a single line of output and went silent. Mythos was a scalpel—precise and dangerously sharp. You could feel the room deciding whether to publish or to lock the lab door.
I watch these product pivots closely, and I want you to understand what Anthropic just did: they released a tamed version of the model everyone whispered about, then kept the raw power behind a velvet rope.
In a Silicon Valley demo room, testers frowned at a single line of output. Anthropic built Mythos and kept it private because it could map holes in security like few AIs before it.
Back in April, Anthropic said it had trained Mythos, a model so skilled at finding cybersecurity gaps that the company feared public release. Access was limited to a handful of organizations through Project Glasswing. Rumors spread: if you heard the name, you wanted to try it; if you ran a security team, you wanted to stop it.
What is Anthropic’s Mythos?
Mythos is Anthropic’s high-capability model class focused on complex reasoning and domain tasks. Anthropic described it as unusually adept at probing systems for vulnerabilities — which, candidly, makes it valuable to both defenders and bad actors. That tension is why Anthropic only let a few partners touch the full-powered version.
On a security call, someone asked whether an AI should spell out an exploit. Anthropic responded by tempering power with guardrails and creating a public-facing model.
Today Anthropic introduced Claude Fable 5, a “Mythos-class model made safe for general use.” The public Fable 5 is presented as exceeding any model Anthropic previously made generally available, while the raw Mythos remains restricted.
Anthropic also launched Mythos 5 for Project Glasswing partners — the same architectural class, but with fewer safeguards. For now, that version stays behind controlled access. Anthropic says broader access will come later after further testing.
Is Fable 5 safe to use?
Fable 5 ships with explicit guardrails. For queries touching sensitive topics — cybersecurity, biological or chemical advice — the model will fall back to an earlier, more conservative model, Opus 4.8, to avoid producing dangerous operational details. The company ran internal red teams and worked with external testers to hunt for jailbreaks and plumbing leaks.
That safety posture means false positives: benign prompts will sometimes be blocked. Anthropic admits it tuned safeguards on the cautious side and plans to relax them over time as they collect feedback. Expect friction if you’re used to a model that always answers — Fable 5 will opt to say less rather than risk harm.
On a billing sheet, the numbers startled a product manager. Fable 5’s price is a clear signal about who Anthropic expects will pay to use it.
Fable 5 and Mythos 5 carry a premium: $10 per million input tokens (€9) and $50 per million output tokens (€46). That’s about twice the cost of the standard Opus 4.8 tier. If you’re an indie developer or a cash-strapped startup, those figures will matter.
The pricing serves two functions: it limits casual experimentation and signals enterprise-level positioning. Still, there’s an aura around Mythos — curiosity can overcome cost. If you’re running threat hunting, legal research, or complex multi-step coding tasks, you might justify the spend; if not, Opus remains the pragmatic choice.
How is Claude Fable 5 different from Mythos?
Think of Fable 5 as a trimmed, safety-first sibling of Mythos. Performance claims suggest Fable 5 outpaces prior publicly available Anthropic models on long, complex tasks — coding, vision, legal and scientific analysis, and tool use. Mythos 5 retains the broader capability set but with fewer filters and restricted availability.
Anthropic’s line-up now gives organizations a menu: Opus 4.8 for conservative, cheaper use; Fable 5 for higher capability with enforced guardrails; Mythos 5 for controlled partner access to the full suite. The move mirrors strategies you’ve seen from other players — think of OpenAI offering tiers of access and Microsoft integrating advanced models into Azure for enterprise customers.
Operationally, the company says both Fable 5 and Mythos 5 were trained to be cautious, accepting some inconvenience to limit real-world harm. Red teams, external security researchers, and partners on Project Glasswing are the crucible where Anthropic will try to reduce false positives without opening a Pandora’s box.
For developers and product leaders, this moment is a trade-off: raw capability versus the moral and regulatory headache of making a powerful tool public. Fable 5 tries to be a seatbelt on a racing car — it keeps the speed but tames the worst consequences.
I’m watching how quickly the safeguards loosen, how many false positives users tolerate, and whether the price keeps Mythos-class models in enterprise hands only. Will Anthropic find the balance between power and restraint, or will the market demand either more access or stricter controls?