Claude Fable 5 explained: What Anthropic's guarded frontier AI model can do

Anthropic's Claude Fable 5 brings Mythos-class AI to public users with safeguards, while the full Mythos 5 model remains restricted to vetted organisations

Anthropic has introduced Claude Fable 5, a new artificial intelligence model that brings its most advanced “Mythos-class” capabilities to general users, while retaining strict controls over how those capabilities can be applied. Alongside it, the company has also released Claude Mythos 5, a separate version of the same system that remains restricted to vetted organisations under its Project Glasswing programme.

 

The dual release marks a shift in how Anthropic is deploying frontier AI systems. Instead of limiting access entirely, the company is separating capability from access, exposing a controlled version publicly while keeping full capabilities within restricted environments.

What is Claude Fable 5

Claude Fable 5 is a Mythos-class model, meaning it operates at the same capability level as Anthropic’s most advanced internal systems. It is designed as a general-purpose AI system capable of handling complex, multi-step tasks across domains such as software engineering, reasoning, and scientific analysis.

 

Unlike earlier Claude models, Fable 5 is built for sustained workflows rather than single prompts. It is designed to plan, execute, and iterate across tasks, making it closer to an agent-style system than a traditional AI chatbot.

What does Claude Fable 5 do

Fable 5 is positioned for high-complexity use cases, including:

• Large-scale software engineering and codebase transformation
• Long-form reasoning and knowledge work
• Vision-based tasks involving structured and unstructured data
• Scientific and technical problem-solving

Anthropic has indicated that performance gains are most visible in longer and more complex tasks, where the model is able to maintain context and execute multi-step workflows.

Mythos Preview, Mythos 5 and Fable 5: What has Anthropic built

Anthropic’s Mythos programme has evolved in three stages, each reflecting a different deployment approach.

 

Claude Mythos Preview was the first version, released under Project Glasswing as a restricted system focused on cybersecurity. It was designed to identify vulnerabilities in software and, in some cases, explore exploit paths across systems. Because of these capabilities, access was limited to vetted organisations, including infrastructure providers and cyberdefence teams.

 

Claude Mythos 5 is the latest version of this system. It builds on Mythos Preview with improved reasoning, coding, and cybersecurity performance. It is a separate model release, not just an update, and continues to operate under restricted access through Glasswing.

 

Claude Fable 5 is the public-facing version of the same underlying system. It shares the same architecture and capability level as Mythos 5, but introduces safeguards that limit how those capabilities can be used.

 

The difference between the three is therefore defined by access and control:

• Mythos Preview: Early, cybersecurity-focused system with strict access controls
• Mythos 5: Updated system with full capabilities, still restricted
• Fable 5: Public version with safeguards and controlled behaviour

How the models differ

Anthropic’s evaluation data shows that Mythos 5 and Fable 5 perform similarly in general-purpose tasks, including coding, reasoning, and knowledge work.

 

In benchmarks such as Agentic coding (SWE-Bench Pro and FrontierCode) and multidisciplinary reasoning (Humanity’s Last Exam), both models outperform earlier Claude versions and competing systems, indicating that their underlying capability is aligned.

 

However, differences emerge in high-risk domains.

 

In cybersecurity evaluations (Firefox, OSS-Fuzz, CyberGym, CyScenarioBench), Mythos systems (Mythos Preview and Mythos 5) show high success rates in completing offensive tasks across environments such as browser exploitation and fuzz testing. These results reflect the model’s ability to identify and act on vulnerabilities.

Fable 5, when safeguards are active, shows a near-complete suppression of these outcomes. In automated adversarial testing, attack success rates drop sharply compared to earlier models, indicating that the system is actively blocking or limiting such behaviour.

 

A similar pattern is visible in biology-related evaluations. Mythos 5 shows higher performance in predicting experimental biological properties compared to Mythos Preview and earlier models, while Fable 5 may restrict or limit responses in these domains depending on the query.

 

This indicates that the difference between Mythos 5 and Fable 5 is not capability, but behaviour.

What safeguards are in place on Claude Fable 5

Claude Fable 5 introduces safeguards designed to address these risks.

 

These include:

• Blocking responses in high-risk domains such as cybersecurity and biology
• Falling back to Claude Opus models for sensitive queries
• Applying conservative filtering to prevent misuse

In adversarial testing, these safeguards significantly reduce the model’s ability to complete offensive cybersecurity tasks. In some benchmark environments, success rates fall to near zero when safeguards are enforced.

 

However, these safeguards are not without trade-offs. They can also block legitimate queries or reduce the quality of responses in sensitive areas, as the system may either refuse to answer or fall back to a less capable model.

Why Mythos systems were restricted in the first place

The decision to restrict Mythos systems is tied directly to their real-world implications.

 

AI models capable of identifying vulnerabilities and generating exploit pathways can compress the time between discovery and exploitation. This is a concern already highlighted by cybersecurity agencies, including frameworks such as those from India’s CERT-In, which note that attackers are increasingly able to act within hours of vulnerability disclosure.

 

Mythos-class systems extend this capability further. They do not just identify issues, but can assist in understanding how those issues can be exploited in practice.

 

This creates a dual-use problem. The same capability that can be used for defence can also be used for attack.

 

This is why Mythos Preview was never released publicly, and why Mythos 5 continues to operate under restricted access.

Do safeguards on Claude Fable 5 address these concerns?

The safeguards directly target the risk of AI-assisted exploit development by reducing the model’s effectiveness in offensive scenarios. This is also clear from the benchmark scores presented by Anthropic.

 

From a policy perspective, this aligns with concerns raised by regulators and cybersecurity agencies about automation in vulnerability exploitation.

 

However, it should be noted that only the outputs are restricted, the underlying capability still exists. This creates a system where risk is managed through control layers rather than removed entirely.

Claude Fable 5 vs GPT models and Gemini

According to Anthropic's benchmark data, Fable 5 outperforms competing models across several agentic and cybersecurity-focused evaluations.

 

For agentic coding, Fable 5 outperforms OpenAI’s GPT 5.5 and Google’s Gemini 3.1 Pro on several benchmarks such as SWE-Bench Pro. It also scores significantly higher than both on Multidisciplinary reasoning benchmarks such as Humanity’s Last Exam.

 

While all three companies are competing across similar areas including agentic workflows, coding, and multimodal systems, their approaches differ in how these capabilities are deployed and integrated.

 

With its newer flagship GPT models, such as GPT 5.5, OpenAI has increasingly focused on building general-purpose agent systems capable of carrying out multi-step tasks across software tools and services.

 

Google, meanwhile, is integrating AI more deeply across its ecosystem with its latest Gemini 3.5 models. Through initiatives such as Gemini Intelligence on Android, Gemini Live, and Gemini-powered Workspace features, the company is focusing on persistent, context-aware AI that can operate across devices, applications, and services.

 

Anthropic's approach differs from both. Rather than positioning Fable 5 primarily as a consumer assistant or ecosystem layer, the company is emphasising high-performance reasoning, coding and scientific work.

 

The release of Mythos 5 alongside Fable 5 also shows that Anthropic is developing specialised frontier models for high-risk domains while maintaining separate safety controls for broader deployment, a step that no competitor has taken yet.

Claude Fable 5 and Mythos 5: Pricing and availability

According to the company, Fable 5 is available globally through the Claude API and consumption-based Enterprise plans from day one. Developers can access the model directly via the API, with pricing set at $10 per million input tokens and $50 per million output tokens. The same pricing applies to Claude Mythos 5.

 

However, access differs significantly between the two models.

 

Claude Mythos 5 remains restricted under Project Glasswing and is currently available only to vetted partners, primarily in cybersecurity. Anthropic also said it plans to extend access to select biology researchers, where safeguards related to biology and chemistry can be lifted in controlled settings.

 

For subscription users, Anthropic is taking a phased approach to Fable 5 rollout.

 

From launch until June 22, the model is included at no additional cost across Pro, Max, Team, and seat-based Enterprise plans. After this period, access will shift to a usage-based model, requiring credits for continued use.

 

The company indicated that this temporary inclusion is tied to capacity constraints, and that it intends to reintroduce Fable 5 as a standard part of subscription plans once sufficient infrastructure is available.

Fable 5 and the IPO race

The release of Claude Fable 5 also comes at a critical moment for Anthropic as it prepares to enter public markets.

 

According to a Reuters report, the company has confidentially filed for a US initial public offering, positioning itself ahead of OpenAI in what is shaping up to be a high-stakes race to define the commercial future of artificial intelligence.

 

The timing is significant.

 

Anthropic last raised $65 billion at a post-money valuation of around $965 billion, more than doubling its valuation from earlier this year and placing it among the most valuable private technology companies globally.

 

This rapid rise reflects investor expectations that frontier AI systems will drive the next phase of growth in the technology sector. At the same time, it raises a key question for the company: how to translate cutting-edge capability into a scalable and defensible business model.

 

This is where Fable 5 fits in.

 

Unlike Mythos Preview, which was restricted under Project Glasswing due to its ability to identify real-world vulnerabilities and perform high-risk tasks, Fable 5 represents a controlled pathway to commercialisation. It brings Mythos-class intelligence into broader use, but with safeguards that limit access to sensitive capabilities.

 

The distinction is likely to be closely watched by investors.

 

As Reuters notes, Anthropic and OpenAI are both moving toward public listings, with the race to go public expected to shape how AI companies present their financial models and long-term viability.

 

In that context, Fable 5 is more than a product launch. It is a signal of how Anthropic intends to balance scale, safety, and revenue as it transitions from a research-driven company to a publicly traded one.