Anthropic Releases Public Fable 5 Model Based on Restricted Mythos AI

The Story

Anthropic has released Claude Fable 5, a public-facing version of its powerful Mythos-class AI model family that had previously been restricted because of cybersecurity risks.

The original Mythos preview was limited to selected partners after the company identified serious dual-use risks, including the model’s ability to help find and exploit software vulnerabilities in critical infrastructure. Fable 5 now brings much of that capability to general users, but with guardrails that route sensitive requests to Claude Opus 4.8 instead of allowing the full model to answer directly.

The release marks a major shift in Anthropic’s model strategy: a safeguarded version of a frontier-level model is now broadly available, while the most sensitive Mythos capabilities remain under controlled access.

The Facts

Access and Pricing

• Claude Fable 5 is generally available through Claude, the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.
• Developers can access the model using the API model ID claude-fable-5.
• Claude Mythos 5 is not generally available and is restricted to approved Project Glasswing partners and selected trusted-access users.
• Project Glasswing expanded in early June 2026 to about 200 organizations in more than 15 countries.
• Both models cost $10 per million input tokens and $50 per million output tokens.
• Claude Mythos Preview cost $25 per million input tokens and $125 per million output tokens.
• Fable 5 and Mythos 5 are priced at less than half the cost of Claude Mythos Preview.
• Fable 5 costs twice as much as Claude Opus 4.8.
• Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost from launch through June 22, 2026.
• From June 23, use on those plans will require usage credits unless Anthropic extends the included access window.

Capabilities

• Fable 5 is Anthropic’s most capable widely released model.
• It is designed for long-horizon agentic work, complex reasoning, coding, research, visual analysis, and enterprise knowledge work.
• Fable 5 and Mythos 5 support a 1 million token context window by default.
• Both models support up to 128,000 output tokens per request.
• Adaptive thinking is always on for both models.
• Raw chain-of-thought content is not returned; developers can receive summarized thinking instead.
• Fable 5 is built for writing and debugging code, answering complex research questions, analyzing images, and handling extended workflows.
• Both Fable 5 and Mythos 5 can work unattended on human instructions for longer periods than any previous Claude models.
• Fable 5 is the public, more heavily safeguarded version of Anthropic’s most advanced Mythos-level technology.

Safeguards

• Fable 5 uses classifiers to detect sensitive requests related to cybersecurity, biology, chemistry, and model distillation.
• When the safeguards trigger, some requests are handled by Claude Opus 4.8 instead of Fable 5.
• These safeguards trigger in less than 5% of sessions on average.
• Users are informed when a request falls back to Opus 4.8.
• Mythos 5 has some safeguards lifted for approved users, depending on the trusted access program.
• Anthropic ran more than 1,000 hours of external red-teaming and a bug bounty program to test whether researchers could bypass the restrictions.
• No universal jailbreak was found during that testing.
• The UK AI Security Institute made progress toward a universal jailbreak during an initial testing window.

Data Retention

• Fable 5, Mythos 5, and future models at similar capability levels require 30-day data retention.
• This data will not be used to train new Claude models.
• The retention policy is intended to support safety monitoring, jailbreak detection, abuse prevention, and false-positive reduction.
• These models are not available under zero data retention.
• The policy overrides previous zero-retention arrangements for some enterprise customers.

What Users Can Use Fable 5 For

Claude Fable 5 is positioned for demanding tasks rather than basic chatbot use. The strongest use cases include:

• complex coding projects
• large codebase analysis
• software migration planning
• document-heavy research
• financial and analytical reasoning
• chart and table interpretation
• visual analysis of screenshots, diagrams, and scientific figures
• long-context writing and synthesis
• agentic workflows that require many steps
• research assistance across technical domains
• fast prototyping of small apps, games, and interactive tools

For general users, the practical value is likely to appear in tasks that normally require several stages: reading large documents, comparing evidence, debugging technical problems, planning complex work, or turning messy inputs into structured outputs.

Vision and Long-Context Work

Fable 5 is also positioned as a stronger vision model. It can extract precise numbers from scientific figures, analyze complex screenshots, and rebuild a web app’s source code from screenshots.

The model also improves on long-running tasks by using persistent memory-like notes. In Anthropic’s testing with the game Slay the Spire, file-based memory improved Fable 5’s performance much more than it improved Opus 4.8.

Mythos 5 and Trusted Access

Claude Mythos 5 is the restricted version of the same underlying model.

It is initially available to existing Claude Mythos Preview users, including Project Glasswing cybersecurity partners. These users can access stronger cybersecurity capabilities with selected safeguards removed.

Project Glasswing partners have included JPMorgan Chase, Apple, Amazon Web Services, Microsoft, Google, and CrowdStrike. Organizations using Mythos have reported finding more than 10,000 critical security flaws in their systems.

Mythos 5 is being deployed in collaboration with the US government. The broader policy context has also changed: the White House has set up an arrangement to test the most powerful models from leading AI companies before release.

Anthropic also plans a trusted access program for biology. This would give selected life science researchers access to Fable 5 with biology and chemistry safeguards removed, while keeping cyber safeguards in place.

This creates a controlled path for organizations that need high-risk capabilities for defensive or scientific work.

Limitations and Practical Risks

The launch comes with several practical limitations.

First, Fable 5 may fall back to Opus 4.8 for some requests. This can affect users working near sensitive areas such as cybersecurity, biology, chemistry, or model distillation.

Second, the subscription rollout is temporary at launch. Users on Pro, Max, Team, and seat-based Enterprise plans get included access only through June 22 unless Anthropic extends the window.

Third, covered-model data retention means these models are not available under zero data retention. This matters for businesses with strict data governance requirements.

Fourth, pricing may matter for heavy users. Fable 5 costs twice as much as Opus 4.8, and intensive coding sessions can consume large token volumes quickly.

Fifth, Fable 5 is deliberately less capable than Mythos 5 in some high-risk areas because the public version routes or restricts sensitive requests.

Sixth, benchmark and early-customer claims remain company-framed. Users should test the model against their own workflows before assuming it will outperform existing models in every task.

Seventh, Mythos Preview was accessed without authorization by a small group of users on the day it was announced. The incident was investigated and contained within a third-party vendor environment.

What to Watch Next

The main questions now are operational rather than theoretical:

• whether Fable 5 can maintain performance under heavy public demand
• whether the safeguard system creates too many false positives for technical users
• how quickly Anthropic expands Mythos 5 trusted access
• how Fable 5 compares with rival frontier models in real-world coding, research, and agent workflows
• whether Anthropic restores Fable 5 as a standard subscription-plan model after the temporary access window

The release also raises a broader industry question: whether future frontier models will be launched as public models with built-in fallback systems, rather than as fully open general-purpose tools.

Why This Matters

Claude Fable 5 shows how frontier AI releases are becoming more powerful and more restricted at the same time. For users, it offers stronger help with coding, research, analysis, prototyping, visual work, and long-context tasks. For the AI industry, it signals a new release pattern: broad access for general tasks, controlled access for high-risk capabilities, higher scrutiny before release, and tighter monitoring for the most advanced models.

This article was drafted with the assistance of generative AI. All facts and details were reviewed and confirmed by an editor prior to publication.