Claude Fable 5: What Anthropic's New Mythos Model Changes for Enterprises

Anthropic just launched Claude Fable 5, its first Mythos-class model made widely available to the public, while Claude Mythos 5 remains reserved for a select group of vetted partners. This rollout of an advanced model, framed by new safeguards, marks a turning point: AI no longer just writes text, but executes long, complex, multimodal tasks with superior autonomy.

Fable 5 vs Mythos 5: Same Base, Two Trajectories

According to Anthropic, Fable 5 and Mythos 5 share the same underlying model, but differ in the level of restrictions applied.

• Claude Mythos 5 is presented as the most capable model for cybersecurity and biology research, available only to a small group of verified partners through trusted access programs (Project Glasswing, security and health clients).
• Claude Fable 5 is the public version of this class: same base capabilities, but with additional safeguards that block or redirect requests in high-risk domains (cyber, biology, chemistry).

Concretely, when a user asks a sensitive question (e.g., help designing malware or toxins), Fable 5 refuses and automatically falls back to Claude Opus 4.8 to produce a more limited, safer response. This architecture allows Anthropic to offer a Mythos-level model to a much broader audience while claiming to limit misuse risks.

A Model Built for Long Tasks and Autonomy

Anthropic describes Claude Fable 5 as "state-of-the-art" in coding, knowledge work, vision, and computer use, with clear advantages over previous models when tasks become long, complex, and context-rich. The company highlights:

• Better performance on very large contexts (millions of tokens), with increased ability to stay coherent and improve its own outputs over time.
• More than 10% performance gains over Claude Opus 4.8 on multiple benchmarks in software engineering, knowledge work, and scientific research.
• Notable gains on vision: extracting numbers from complex scientific figures, reconstructing code or interfaces from screenshots, analyzing visual documents.

The direction is clear: Fable 5 targets engineering workflows, expert analysis, and long multimodal tasks, where older models tended to "drop out" or contradict themselves.

Real-World Examples: Stripe, GitHub, AWS

Stripe: Migrating 50 Million Lines of Ruby Code

Anthropic reports that Stripe used Claude Fable 5 to execute a large-scale migration in a Ruby codebase of approximately 50 million lines. This project, which would have required months of work from a dedicated team, was reportedly compressed into a single day thanks to the model, according to reported tests: massive refactoring, consistent updates, and change verification.

This case well illustrates Fable 5's promise on long engineering tasks: where older models would need many round-trips and very fine manual chunking, Fable 5 is designed to follow a larger plan, keep the thread, and reduce human friction.

GitHub Copilot: A Mythos Model for Code

GitHub announced the integration of Claude Fable 5 into GitHub Copilot, alongside other models like Claude Opus 4.8 and OpenAI models. According to GitHub, Fable 5 is specifically positioned for autonomous coding workflows and long knowledge tasks, with fewer tool calls and lower token consumption than Opus models on their internal benchmarks.

GitHub notes, however, that Fable 5 enforces 30-day data retention to allow Anthropic to run its security classifiers, while other Claude models in Copilot remain in "Zero Data Retention". Administrators must explicitly activate this option in settings, showing that governance and compliance concerns are real for enterprises.

AWS: Mythos Accessible via the Cloud

AWS also communicated about Claude Fable 5 availability through its services, presenting it as a way to benefit from Mythos capabilities with built-in safeguards, particularly for advanced engineering and knowledge work workloads. The pitch: allow cloud customers to leverage a very powerful model without assuming the same risks as direct Mythos 5 access.

Scientific Performance and R&D: Hypotheses, Biology, Cybersecurity

Anthropic and several external analyses also highlight the role of Fable 5 and Mythos 5 in scientific domains:

• Internal protein-design experts reportedly observed a 10x acceleration on certain process steps thanks to the Mythos model, with new molecular hypotheses preferred roughly 80% of the time against outputs from older Opus models in blind tests.
• The Mythos 5 version, not accessible to the general public, is explicitly described as "state-of-the-art for cybersecurity, biology research, and health", explaining its very limited, controlled distribution.

These uses are precisely what motivates Anthropic's caution: a model's ability to find software vulnerabilities or propose pathways for pathogens can serve to secure… or to attack. This is why safeguards sit at the heart of the Fable/Mythos architecture.

Security and Data Retention: An Assumed Trade-off

To make a model of this power accessible more broadly, Anthropic relies on a combination of security classifiers and limited data retention:

• All requests to Claude Fable 5 are submitted to classifiers that detect potentially dangerous uses and block or redirect responses by domain (cyber, bio, chemistry).
• The company explains that less than 5% of sessions trigger these safeguards, meaning the vast majority of "normal" uses are unaffected.
• To operate these protections, Anthropic retains prompts and responses for up to 30 days, solely for abuse detection, before deletion. These data reportedly are not used for model training.

This policy differs from "Zero Data Retention" models and raises important questions for organizations subject to strong confidentiality, sovereignty, or trade-secret requirements. In the GitHub Copilot example, activation of Fable 5 is left to administrator discretion, with explicit warnings about retention.

Pricing and Positioning: A Clearly Premium Model

In pricing terms, Claude Fable 5 sits in the high-end model category:

• $10 per million input tokens and $50 per million output tokens on the API—roughly double Claude Opus 4.8 rates (5 / 25).
• The model is free on Pro, Max, Team, and Enterprise plans from June 9–22, 2026, then switches to credit-based usage from June 23 onward.

Several observers note this premium only justifies itself for use cases where output quality and depth significantly offset token cost: critical code migrations, complex analyses, highly sensitive documents, high-value autonomous agents. For simpler tasks (FAQs, short summaries, basic text generation), more economical models remain viable.

When to Consider Claude Fable 5

For enterprises, Fable 5's value doesn't lie in being "stronger" in general, but in the types of problems it handles better than the previous generation:

• Long tasks with substantial context to maintain.
• Environments requiring manipulation of code, documents, spreadsheets, or screenshots.
• Scenarios where human oversight of each step is costly (complex analysis, massive refactoring, multi-source detailed synthesis).

The Stripe, GitHub, and AWS examples illustrate a key point: the most transformative use cases are those where the model receives a hard, well-specified mission and can work "in depth" rather than produce a quick, superficial response.

Key Takeaways

• Claude Fable 5 is the first Mythos-class model accessible to the general public, with specialized safeguards on sensitive domains.
• It targets engineering workflows, knowledge work, and long-vision tasks, where it significantly outperforms Opus 4.x models across multiple benchmarks.
• Real examples (Stripe, GitHub Copilot, AWS customers) show genuine productivity gains on massive codebases and complex tasks.
• In exchange, it enforces 30-day data retention and premium pricing ($10 / $50 per million tokens), making it a tool reserved for high-value cases and requiring careful governance framing.

For product, data, security, and IT teams, the challenge is not adopting Fable 5 for everything, but identifying it as an ultra-capable building block in a model portfolio—activated where its premium cost translates into real gains in quality, speed, or security.