Claude Fable 5: Anthropic’s First Public Mythos-Class Model

With Claude Fable 5, Anthropic is offering the public its most capable model yet — and openly admitting it isn't safe enough to hand over without guardrails that can override the user's choice. It's a bet that heavy safety machinery is the price of frontier access, and a preview of the access tiers that may define the next phase of AI.

What Is Claude Fable 5?

Anthropic just released their new Fable 5 model – their first publicly available model that is "Mythos-class", that is to say, a completely different level of model than all their previous ones. Anthropic first announced the existence of this class of model back in April and in a break with previous practices, limited access to this model to a consortium of vetted partners known as Project Glasswing. According to Anthropic, Mythos-class models are so powerful that ordinary users are not able to even use it for cybersecurity or certain biology tasks, because Anthropic says they can't guarantee that malicious actors won't make use of it in harmful ways.

Claude Fable 5 vs Mythos 5: What's the Difference?

The key story with Claude Fable 5 is Anthropic has finally taken the model class it previously kept behind closed doors and made a version available to the public. But it has done so with a new kind of compromise. The raw model, Claude Mythos 5, is still restricted to vetted partners. The public version, Claude Fable 5, runs on the same underlying model but adds safeguards that detect certain risky requests and route them away from Fable.

This two-tier structure is not like previous model launches. Fable and Mythos are not really “small model versus big model” (for example Sonnet versus Opus), but they are more like “public model with guardrails (Fable 5)” versus “restricted model with some guardrails lifted (Mythos 5).” Mythos 5 is reserved for Project Glasswing partners such as cyberdefenders and critical infrastructure providers. Fable 5 is the version normal users and developers can access. That may foreshadow where frontier AI is going: not one model for everyone, but different access tiers depending on who you are, what you are doing, and how much risk the lab is willing to tolerate.

Capabilities and Performance of Claude Fable 5

Anthropic’s own system card (a technical report that documents an AI model’s capabilities, limitations, safety evaluations, and known risks) is a sign of how seriously the company wants this release to be taken. It is 319 pages long, which is a new benchmark for Anthropic (the previous Opus 4.8 model card clocked in at 246 pages). One can see this 300+ page document as part safety case, part benchmark dump, part risk disclosure, and part argument for why Anthropic thinks it is acceptable to release a Mythos-class model at all.

On the new model’s capability, Anthropic claims that Fable is much stronger than previous public Claude models. All tested benchmarks show it is state-of-the-art, especially in software engineering, long-horizon agentic tasks, knowledge work, vision, and scientific reasoning. Anthropic specifically singles out its performance on advanced scientific research tasks like, for example, helping design new biological structures, such as virus-like delivery systems, and predict whether they would function as intended. Anthropic claims that the Mythos-class models were able to match or exceed strong human and specialist-tool baselines in some tests.

Pricing, Access, and Availability

Anthropic’s pricing scheme for Fable is also notable. The recent trend in frontier models (not just Anthropic’s but those of their frontier lab peers) has been that a new release is either better or cheaper. Fable is, in a sense, both. Sort of. It is not cheap in absolute terms: at $10 per million input tokens and $50 per million output tokens, it costs twice as much as Opus 4.8, the previous Anthropic state-of-the-art model. However, this pricing is less than half of Mythos Preview, Anthropic’s earlier restricted Mythos-class model. So Anthropic is pushing the frontier forward while making a once-restricted tier cheaper to reach.

But the catch here is (and this is something new for Claude plan subscribers), Fable 5 is only accessible in Pro, Max, Team, and seat-based Enterprise plans until June 22. Afterwards, use of Fable 5 requires pay-as-you-go usage credits, unless Anthropic decides to extend the window. That makes the launch feel partly like a public preview and partly like a test of the public’s willingness to pay. Everyone gets to taste the frontier, but continued access will be metered.

Safety Guardrails and the Alignment Debate

On the alignment question, the release becomes more complicated. Anthropic says Fable is among its most aligned models, but the actual safety strategy is not just “the model is safe now.” It is “the model is safe enough after we put in ways for it to refuse or redirect an answer”. If Fable detects certain cybersecurity, biology, chemistry, or distillation-related prompts, it does not answer as Fable; instead it defaults to answering as a less capable model, like Opus.

What Claude Fable 5 Means for the Future of AI

For some users, this is controversial because it means they are not always getting the model they selected. A cybersecurity researcher, biologist, or advanced developer may hit a safeguard even on benign work. More controversially, Anthropic has also added hidden interventions for frontier AI development tasks. These do not necessarily show up as a clear refusal to answer. Instead, the model itself limits its own effectiveness on performing certain AI training or programming tasks. Critics see this “feature” of the model limiting its own usefulness without clearly telling the user as opaque and paternalistic.

For others, Anthropic has not gone far enough. The system card contains uncomfortable results for those who are willing to dive into it. For example, in Vending-Bench Arena (an existing benchmark that evaluates how effectively AI agents can run a simulated vending-machine business over extended periods), Fable 5 was the only tested model to initiate unethical behavior like price fixing. It even rationalized behavior it appeared to recognize as unethical, reframing the price-fixing as “market stabilization.” This is exactly the kind of unsettling behavior that makes people nervous as models become more autonomous.

So, the news here is that Fable 5 is not just another frontier model release, an occasion that happens now more and more often, for good or ill. One can see it as a preview of the next phase of AI deployment with new issues highlighted: partial public access, heavy safety infrastructure, usage-based economics, and growing fights over who gets the real model. Anthropic’s argument is that this is the responsible way to release something this powerful. The counterargument is that if the model needs this many caveats, maybe the public release is less settled than the launch suggests.