Anthropic launches Claude Fable 5 and limited access to Mythos 5

Anthropic has announced Claude Fable 5 and Claude Mythos 5 today. That matters because the Mythos name has been circulating for weeks as the rumored label for an extremely powerful Claude model focused on cybersecurity. The official launch is more nuanced: Fable 5 is the Mythos-class model becoming broadly available, while Mythos 5 itself remains limited to a small group of cyber defenders and infrastructure partners.

According to Anthropic, Claude Fable 5 is the most powerful model the company has made widely available so far. It belongs to the new Mythos class, a tier above the Opus models. Anthropic says Fable 5 is stronger in software engineering, knowledge work, vision, long-running tasks, memory and scientific research, with its advantage growing as tasks become longer and more complex.

Claude Mythos 5 is, according to Anthropic, the same underlying model as Fable 5, but with some safeguards lifted. That makes Mythos 5 more directly useful in sensitive areas such as cybersecurity. For that reason, it is not being released broadly. Anthropic says Mythos 5 will initially be deployed through Project Glasswing, in collaboration with the U.S. Government, as an upgrade for users who already had access to Claude Mythos Preview.

The core of the launch is this: Anthropic is bringing Mythos-level capability to the public, but not in its most open form. Fable 5 comes with additional safeguards. For higher-risk requests involving cyber, biology, chemistry and model distillation, Axios and Anthropic report that the system can route users to the less capable Claude Opus 4.8 instead. The idea is that ordinary users get better reasoning power while dangerous uses are slowed down.

That is a notable choice. Many AI launches are framed around speed, benchmarks and price. This one is just as much about access control. Anthropic is effectively saying that this model class is powerful enough to help with software, science and security, but also powerful enough to be misused. So the public gets Fable 5 with guardrails, while Mythos 5 expands through a trusted-access program.

The background explains the caution. In Project Glasswing, Anthropic used Mythos Preview to examine critical software. The company previously wrote that Mythos Preview found vulnerabilities in major software projects and that partners sharply increased their bug-finding speed. Anthropic said partners collectively found more than ten thousand vulnerabilities after one month. Mozilla found and fixed 271 vulnerabilities in Firefox 150 while testing Mythos Preview.

That shows the dual-use nature of the model. For defenders, this kind of AI can be extremely valuable. If AI finds vulnerabilities faster, companies and open-source projects can patch earlier. But the same capability can also help attackers. A model that is good enough to find vulnerabilities and reason through exploits is not just a better chatbot. It is a tool that can shift the balance of cybersecurity.

For developers, Fable 5 matters because Anthropic says it can work autonomously for longer than previous Claude models. The company points to codebase migrations, more complex agentic coding and better performance on long tasks. The promise is that AI will not only write isolated pieces of code, but plan, execute, check and revise larger software assignments.

The price is high. Anthropic lists both Fable 5 and Mythos 5 at $10 per million input tokens and $50 per million output tokens. That is more expensive than many existing models, but Anthropic positions it as a model that can be cheaper per task if it needs fewer steps, fewer corrections and less human intervention.

My conclusion: the Mythos launch is not simply “Claude Mythos for everyone.” It is a two-part move. Anthropic is making Mythos-class capability broadly available for the first time through Fable 5, while keeping the more open Mythos 5 access behind a controlled gate. That makes this launch important not only because of performance, but because it shows how AI companies are starting to handle models that are both productive and risky.