Anthropic has released Claude Fable 5, its most capable publicly available AI model to date, while simultaneously introducing a more permissive version called Claude Mythos 5 that will remain restricted to a select group of trusted partners.

The dual launch reflects a growing reality facing frontier AI developers: the most advanced models are becoming powerful enough that companies increasingly feel compelled to limit access to certain capabilities, even as they continue pushing performance forward.

According to Anthropic, Claude Fable 5 is the company’s first publicly available Mythos-class model and represents a major leap in software engineering, reasoning, scientific research, and long-horizon task execution. The model is available immediately to users through Anthropic’s Claude platforms.

But the more consequential announcement may be what Anthropic chose not to release broadly.

Alongside Fable 5, the company unveiled Claude Mythos 5, a restricted variant built on the same underlying foundation model. Unlike Fable 5, Mythos 5 will initially be available only through Project Glasswing, Anthropic’s trusted-access initiative for cybersecurity defenders, critical infrastructure operators, and selected institutional partners.

The distinction highlights Anthropic’s evolving approach to AI deployment.

Rather than treating AI access as a simple choice between public and private, the company is creating different levels of access based on the potential risks associated with increasingly capable systems.

A New Tier of AI Access

Anthropic says Fable 5 includes safeguards designed to detect and redirect a narrow category of high-risk requests. When triggered, those requests are routed to Claude Opus 4.8, a model Anthropic considers safer for handling sensitive interactions.

The company noted that these safeguards affect fewer than five percent of user sessions, though it intentionally tuned them conservatively to prioritize safety.

Mythos 5, by contrast, removes some of those restrictions for approved organizations operating in defensive and research-oriented environments.

Anthropic argues that this controlled-access model allows critical sectors to benefit from frontier AI capabilities while reducing the risk of broader misuse.

The approach reflects concerns that advanced AI systems are becoming increasingly dual-use in nature—capable of accelerating both beneficial and harmful activities depending on how they are deployed.

Among the areas Anthropic specifically highlighted were cybersecurity, biological and chemical research, and model distillation, where powerful AI systems could potentially be used to extract, replicate, or enhance capabilities in unintended ways.

Performance Gains Across Complex Tasks

While Anthropic devoted considerable attention to safety, the company also emphasized the model’s technical advances.

Fable 5 reportedly delivers state-of-the-art performance across a wide range of benchmarks, with particularly strong gains in software engineering and complex reasoning tasks.

According to Anthropic, the model’s advantages become more pronounced as tasks grow longer and more difficult. The company highlighted examples ranging from autonomous coding projects and scientific research to advanced vision tasks requiring multi-step reasoning.

Among the demonstrations cited was the successful migration of a 50-million-line Ruby codebase in a single day—a task that could otherwise require months of engineering effort. Anthropic also pointed to progress in molecular biology research, autonomous simulations, 3D modeling, and game-playing tasks performed using raw screenshots and limited tools.

The message is clear: AI systems are no longer merely answering questions. They are increasingly being deployed to execute complex workflows that previously required teams of skilled professionals.

Capability Growth Meets Safety Concerns

The launch comes just days after Anthropic publicly warned that AI development may be approaching a critical inflection point.

In a recent paper, the company argued that advances in AI-assisted software development are accelerating rapidly and could eventually enable recursive self-improvement—the ability of AI systems to contribute directly to the development of more capable successor systems.

Anthropic stopped short of claiming that such capabilities already exist. However, it warned that progress may be occurring faster than many institutions are prepared for and called for discussions around potential mechanisms for slowing frontier AI development if safety measures fail to keep pace.

Against that backdrop, the Fable 5 and Mythos 5 split appears less like a product decision and more like a governance strategy.

Anthropic is effectively acknowledging that some AI capabilities may be too powerful for unrestricted deployment while still arguing that those capabilities can be valuable when placed in the hands of trusted organizations.

A Glimpse of AI’s Next Debate

For years, the AI industry has debated whether advanced models should be open or closed.

Anthropic’s latest release suggests the next phase of that conversation may focus on a different question entirely: who should have access to the most powerful capabilities, and under what conditions?

The emergence of tiered-access models could become increasingly common as frontier systems continue to improve. Governments, critical infrastructure providers, cybersecurity firms, and research institutions may gain access to capabilities unavailable to the broader public, creating a new category of trusted AI users.

Whether that approach proves effective remains to be seen.

What is clear is that the industry’s leading developers are no longer treating AI safety as a challenge that can be addressed solely through content filters and usage policies. Increasingly, they are experimenting with structural controls over who gets access to the most advanced systems in the first place.

With Claude Fable 5 and Mythos 5, Anthropic is offering a glimpse of what that future might look like: powerful public AI for everyone, and even more powerful AI reserved for a select few.

‍