Anthropic Splits Claude Models to Balance Power and Security

Anthropic has released two versions of its latest large language model, Claude 5, to address growing concerns about AI misuse. The company is offering Claude Mythos 5 to a select group of cybersecurity experts and infrastructure providers, while Claude Fable 5 is available to the broader public with stricter limitations. This dual-release strategy aims to harness the models’ capabilities while mitigating risks associated with their power.

Why Anthropic Split Its Models?

Anthropic’s decision reflects a broader challenge in AI development: how to deploy highly capable systems without enabling malicious use. The Mythos 5 version, designed for trusted partners, has already uncovered over 10,000 critical security flaws in systems tested under Project Glasswing, an initiative launched in April 2024. However, the same tools that identify vulnerabilities could also be exploited by attackers to breach systems, according to internal assessments cited by Wired.

“We’re trying to make improvements in a way that’s beneficial, even if we don’t have the perfect solution for every use case,” said Diane Penn, Anthropic’s head of product management. The company emphasizes that Mythos 5 remains restricted to “a small group of cyberdefenders and infrastructure providers,” with access determined by a need-to-know basis.

How Does Fable 5 Differ from Mythos 5?

Fable 5, the general-purpose version, operates on the same underlying model as Mythos 5 but includes strict safeguards. It blocks queries related to cybersecurity, biology, and chemistry, redirecting them to an older model, Claude Opus 4.8. The system also detects attempts to “distill” its outputs into smaller models, a practice that could enable third parties to replicate its capabilities without oversight.

Anthropic acknowledges that these measures may inadvertently block harmless queries. “We’re erring on the side of over-blocking for now,” Penn said. The company plans to refine its classifiers over time but argues that caution is necessary to justify a wider release.

What Are the Implications for AI Safety?

The split highlights a tension in the AI industry between innovation and control. Anthropic’s pricing reflects the models’ power: $10 per million input tokens and $50 per million output tokens for both versions, roughly double the cost of its earlier models. This pricing strategy positions Mythos 5 and Fable 5 as specialized tools rather than general-purpose AI.

Competitors like OpenAI are taking similar approaches. The company has introduced an advanced cybersecurity model for select partners and formed a working group akin to Anthropic’s Project Glasswing. Both firms face pressure to demonstrate safety while preparing for potential IPOs, according to BBC reports.

How Will This Shape the Future of AI?

Anthropic’s approach underscores a growing debate about AI governance. Jack Clark, co-founder of the company, has warned that the industry lacks “a brake pedal” to slow down development. “You want the option to be able to take your foot off the gas and put your foot on the brake,” he said. “Right now, it’s like the AI industry has a gas pedal, but it doesn’t have a brake pedal.”

As AI systems become more autonomous, their potential to perform unattended tasks—such as software development or data analysis—raises new ethical questions. The success of models like Fable 5 will depend on balancing accessibility with safeguards, a challenge that could shape the next phase of AI regulation.