Claude Fable 5: Anthropic releases Mythos, but with the safety brakes on

CursorBench points in the same direction. Fable 5 appears as one of the strongest models for long and difficult coding tasks, with a strong balance between performance and cost per task. Cursor even describes it as opening up a new category of “long horizon” problems that earlier models could not reliably handle.

Fewer prompts, more autonomy

One of the most interesting reactions came from Andrej Karpathy, who described Fable 5 as a real qualitative jump, especially during long problem-solving sessions. His point is not just that the model performs better on benchmarks. It is that it makes users want to give it much more ambitious tasks.

That is probably where Fable 5 becomes interesting for companies. The gain does not only come from smarter answers, but from the fact that the model seems to require less micromanagement. Fewer follow-ups, fewer corrections, fewer intermediate explanations. If the model understands the broader objective more reliably, the user can move from assisted executor to supervisor.

This also makes these models harder to evaluate. A traditional benchmark measures an answer. A company wants to know whether a model can stay on a long task, recover from errors, keep context, understand an architecture and produce something usable.

Vision, long memory and Pokémon FireRed

Anthropic also highlights Fable 5’s multimodal capabilities. The most viral example is probably Pokémon FireRed: the model reportedly finished the game using only vision, while keeping memory across a long context.

Taken alone, the example may sound anecdotal. But it illustrates something more serious: a model’s ability to observe an environment, remember what it has already done, plan several steps ahead and keep progressing without constant human instructions.

That is exactly what today’s assistants often lack. They are good at one-off requests, but struggle when a task becomes long, visual, interactive or filled with small sequential decisions.

Science and finance as strategic showcases

Anthropic is also positioning Fable 5 and Mythos 5 around more specialized use cases. In biology, the company says Mythos 5 accelerated parts of the drug design process, with workflows capable of choosing binding sites, running bioinformatics tools and recovering from failures.

That example matters, but it also reveals the limits of Fable 5. The most sensitive capabilities seem to be precisely the ones Anthropic does not want to make fully public.

In finance, Anthropic highlights performance on document analysis, financial research benchmarks and complex information extraction. For companies, this is probably one of the most immediate use cases: reading reports, comparing figures, extracting information from long documents, producing structured analysis and automating knowledge work.

More expensive, but potentially more profitable

Fable 5 is more expensive than previous Claude models. Its announced price is $10 per million input tokens and $50 per million output tokens. For eligible Claude Pro, Max, Team and Enterprise users, Anthropic is offering an included access window before usage switches to credits.

At first glance, that is expensive. But raw cost per token is not necessarily the right metric. If Fable 5 solves a task with fewer back-and-forths, less repeated context and fewer human corrections, the real cost can become more attractive.

The most advanced models are not meant to replace fast and cheap models for every small request. They are designed for problems where failure is more expensive than inference: code migration, financial analysis, scientific research, document review, internal automation and software agents. The right question is no longer “how much does an answer cost?”, but “how much does a correctly completed task cost?”

The real question: who gets access to full power?

Fable 5 points to a trend that will probably accelerate: the best models will not necessarily be available in the same way to every user.

The public will get a protected version. Critical companies may get a more open one. Validated researchers will access specific capabilities. Users considered risky may be limited, routed or blocked. This is no longer only a question of price or subscription tier. It is a question of status, trust and control.

That logic can be defended. If a model can genuinely help discover vulnerabilities, automate sensitive research or manipulate complex systems, publishing it without filters would be irresponsible. But it also raises a question of transparency. Users will need to understand when they are using the full model, when they are using a restricted version and why certain capabilities are denied to them.

Fable 5 is therefore not just a more powerful Claude. It is a signal of how frontier models may be distributed in the future: not as identical tools for everyone, but as systems adjusted according to the level of trust granted to each user. The public gets Fable, validated actors get Mythos, and Anthropic is testing a new boundary between the two: giving access to power, without fully giving up control.