Anthropic released Claude Fable 5 on June 9, and the release itself is a story about security architecture as much as model capability. Fable 5 is the first publicly available model in what Anthropic calls its Mythos class, a tier that now sits above Opus in the company’s lineup.
If the Mythos name sounds familiar, it should. Mythos Preview launched in April to a small set of partners, and the reason for the restricted rollout was blunt: the model proved unusually good at finding and exploiting vulnerabilities across every major operating system and browser, despite not being built for cybersecurity work. That capability is what kept it locked behind a trusted-access program for two months. Fable 5 is Anthropic’s answer to the question of how you ship that class of model to everyone else.
The same model, two doors#
Fable 5 and Mythos 5 share the same underlying model. The difference is access control.
Mythos 5 went out the same day to organizations already approved under the earlier preview, including the cybersecurity partners in Project Glasswing, Anthropic’s collaboration with major cloud, software, and financial firms that used early Mythos access to find and fix vulnerabilities in their own products. Anthropic expanded that circle to hundreds of organizations across 15 countries in the week before launch, with a stated focus on critical infrastructure operators, and says a more systematic trusted-access program is coming.
Fable 5 is the general-availability version. It carries additional safeguards: when a request lands in one of four designated high-risk areas, namely cybersecurity, biology, chemistry, or model distillation, Fable 5 doesn’t answer. The request falls back to Claude Opus 4.8, which produces the response instead. Anthropic reports the fallback fires in fewer than 5 percent of sessions, so the large majority of users get the full Mythos-class model the large majority of the time.
What it’s actually good at#
Anthropic’s claim is that Fable 5’s capabilities exceed any model the company has ever made generally available, with state-of-the-art results on nearly all tested benchmarks. The areas it highlights are software engineering, knowledge work, vision, and scientific research, and the company says the model’s lead over its other models grows as tasks get longer and more complex.
Some concrete signals from launch coverage:
On some benchmarks, Fable 5 scored more than 10 percent higher than Claude Opus 4.8, which itself shipped only weeks earlier. Analytics company Hex reported it was the first model to clear 90 percent on its core analytics benchmark. And in early customer testing, Stripe reported the model completed a Ruby codebase migration in days that had been estimated at more than two months of team effort.
That last data point is the one that matters for anyone planning engineering capacity. The pitch for this model class isn’t better chat responses. It’s sustained, multi-step autonomous work: long migrations, large refactors, extended research tasks. The economics of that pitch are visible in the pricing, which assumes the model replaces project time, not message volume.
On the restricted side, Anthropic says Mythos 5 has shown meaningful acceleration in parts of drug-design workflows for approved biology researchers. The same capability profile that makes the model useful for vulnerability discovery makes it useful for legitimate research, which is exactly why the access tiers exist.
The jailbreak question#
Capability gating only matters if it holds. Anthropic says it ran an external bug bounty that produced no universal jailbreak across more than 1,000 hours of testing, then brought in external red-teaming organizations that also failed to find one.
That’s a stronger pre-release validation posture than most model launches get, and it deserves credit. It’s also not a closed question. The absence of a universal jailbreak in testing isn’t the absence of bypass techniques, and a sub-5-percent trigger rate means the classifier is making judgment calls at the boundary of those four domains constantly. Anthropic acknowledges novel attacks remain possible, and the company has added a 30-day data retention policy on all Fable 5 and Mythos 5 traffic for safety monitoring, with a commitment that the data won’t be used for training.
That retention policy is a real consideration for organizations with data-handling obligations. Thirty days of mandatory retention for safety review is a different posture than zero-retention API arrangements some enterprises have negotiated, and it should be evaluated like any other processor data-handling term.
Pricing and access#
Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, twice the cost of Opus 4.8 and less than half what Mythos Preview cost its partners. It’s available through the API and consumption-based enterprise plans now.
The two-tier model is the precedent#
The benchmarks will be old news in a quarter. The release architecture won’t be.
Fable 5 establishes a pattern: frontier capability ships publicly with specific dangerous domains gated at runtime, while the ungated version goes to a vetted population under a trusted-access program. That’s a familiar structure to anyone who has worked with export controls or dual-use research, applied for the first time at this scale to a commercial AI product.
It also creates a planning assumption security teams should write down now. Mythos-class capability exists, it’s in production, and several hundred organizations have the unrestricted version. The defensive applications are real, and Project Glasswing’s vulnerability-hunting work is the proof case. But capability diffusion only moves in one direction. Open-weight models trail frontier models by months, not years, and the vulnerability-discovery capability that justified Mythos’s restricted rollout is the same capability defenders should assume adversaries are working toward independently.
The practical takeaway isn’t alarm. It’s that patch latency, exploit-prediction assumptions, and vulnerability management SLAs were calibrated against human-speed offensive research. Those assumptions now have a shelf life, and releases like this one are the timer.
The release also landed days after Anthropic publicly urged frontier labs to coordinate on slowing development ahead of recursive self-improvement risks, a juxtaposition the company will keep getting asked about. Shipping the most capable public model in the industry while calling for a collective brake is a tension you can read as hypocrisy or as exactly what a race-to-the-top strategy looks like in practice. Probably both readings will show up in your feed this week.
References#
- Anthropic: Claude Fable 5 and Mythos 5 announcement
- TechCrunch: Anthropic’s Claude Fable 5 is a version of Mythos the public can access today
- IT Pro: Anthropic just launched Claude Fable 5, its first Mythos-class AI model
