Anthropic launches Claude Fable 5 and Mythos 5

Anthropic has launched Claude Fable 5, its most capable AI model available to the public, alongside Claude Mythos 5, a less restricted version initially reserved for cyberdefenders and select strategic partners.

The difference between the two is not the base model. Fable 5 and Mythos 5 share the same technology, but Fable includes additional systems that block or redirect certain requests related to cybersecurity, biology, chemistry, and extracting capabilities to train other models.

What Fable 5 can do

Anthropic says Fable 5 outperforms its previous models in programming, document analysis, image interpretation, scientific research, and long-running tasks that require maintaining context over extended periods.

An example cited by Stripe shows the type of work it aims to automate: Fable 5 migrated a Ruby codebase with 50 million lines in one day, a task that the company says would have occupied a team for more than two months.

In practice, this means you can ask it to work on a complex project for longer instead of limiting yourself to isolated questions. For example, it can analyze an application, propose changes, write the code, test it, and fix errors with less human intervention.

Anthropic also highlights improvements in:

• Autonomous programming, with fewer turns and a greater ability to complete long tasks.
• Financial and document analysis, including tables, charts, and lengthy documents.
• Vision, such as reading figures in scientific charts or reconstructing an application from screenshots.
• Persistent memory, through files or notes that the model can consult during a long-running task.
• Scientific research, especially in molecular biology, genomics, and protein design.

The results come from internal evaluations and tests conducted by partner companies, so they should be understood as data reported by Anthropic, not as independent validation of all its claims.

Mythos 5 is reserved for high-risk uses

Claude Mythos 5 is initially aimed at cybersecurity organizations, critical infrastructure providers, and US government partners within Project Glasswing. Anthropic describes it as its strongest cybersecurity model to date.

The company plans to expand access through a trusted-access program. It is also preparing a dedicated program for biomedical researchers, with its biology and chemistry restrictions adjusted while keeping protections related to cyberattacks in place.

Anthropic says Mythos 5 accelerated some internal drug-design processes by approximately ten times. In one test, the model performed tasks normally handled by a scientist, such as selecting binding sites, using protein-design tools, and recovering from errors.

The company also says that nine of the 14 protein targets studied produced candidates that are now under investigation. These are preliminary results, not drugs ready for use.

Fable 5's safety limits

Requests that the safety systems consider particularly sensitive are not answered by Fable 5. Instead, the conversation is passed to Claude Opus 4.8, a less capable model in those areas.

This can happen with requests involving:

• Exploit development and cyberattacks.
• Biology and chemistry with potential harmful uses.
• Attempts to extract the model's capabilities to train competing systems.

Anthropic says these redirects are activated in less than 5% of sessions on average. The trade-off is that some legitimate requests may also be blocked, especially in scientific research and cybersecurity.

The company says it subjected the protections to more than 1,000 hours of external testing and that no universal method for bypassing them was found. Even so, it acknowledges that no system of this kind can guarantee perfect protection and that it will continue adjusting the filters.

Pricing, privacy, and availability

Both models cost 10 dollars per million input tokens and 50 dollars per million output tokens. A token is a unit of text that the model processes, so the cost depends on the length of your conversations and responses.

Fable 5 is available through the Claude API and is gradually being added to subscription plans. Until June 22, it will be included at no additional cost in Pro, Max, Team, and seat-based Enterprise plans. After that, it will require usage credits, at least temporarily, depending on available capacity.

For enterprise customers, Anthropic will apply a mandatory 30-day retention period to all traffic involving Mythos-class models. The company says this data will not be used to train new models or for purposes unrelated to security, and that it will log human access before deleting the data in almost all cases.

For you, the main change is that Claude can take on longer, more complex work, especially in programming, analysis, and documents. But it also comes with a clearer boundary: the closer a task is to offensive cybersecurity or high-risk biology, the more likely Fable 5 is to redirect it to a less capable model. What remains to be seen is whether Anthropic can reduce those false positives without opening the door to malicious use of increasingly autonomous capabilities.