It can stay with complex tasks and hard problems for days, handling ambitious work across coding, research, analysis, finance, law, and beyond.
Claude Fable 5, Anthropic’s new “Mythos-class” model, is specifically engineered to handle highly complex, multi-step work that previously required constant human supervision. Its suitability for ambitious projects stems from several key factors:
- Long-Horizon Autonomy: The model is designed for persistent, multi-day operations. It can stay with complex problems and operate for days without intervention (1:22 – 1:36), making it ideal for tasks like large-scale code migrations or deep research projects.
- Self-Correction and Agentic Capabilities: Fable 5 features built-in verification loops, enabling it to test its own code, self-correct, and manage sub-agents to complete intricate, multi-step projects.
- Advanced Reasoning & Coding: It demonstrates superior capability in complex coding, data-driven agentic workflows, and 3D design, allowing it to take on challenging projects across fields like finance, law, and economics.
- Vision-Based Analysis: The model can interpret and analyze complex technical figures or screenshots, and even rebuild applications based on visual input.
- Enterprise-Grade Safety: While powerful, it includes robust, automatic safeguards. If a request touches on high-risk areas like cyber security or biology, the system is designed to intelligently redirect those specific parts of the request to Opus 4.8, ensuring safety without compromising general productivity (0:52 – 1:13).
Claude Fable 5 is Anthropic’s first generally available “Mythos-class” intelligence model, built specifically for highly autonomous, multi-day engineering and complex knowledge tasks. Released alongside the highly restricted Claude Mythos 5, Fable 5 introduces a new era of AI focused on handling long-horizon responsibilities independently rather than executing isolated, short-term tasks. [1, 2, 3]
Core Model Specifications
- Context Window: 1 million tokens default input capacity.
- Output Limit: Up to 128,000 output tokens per single request.
- Pricing: $10 per million input tokens and $50 per million output tokens.
- Adaptive Thinking: Forced natively; “raw” unsummarized thinking content is permanently omitted to protect backend reasoning architecture. [2]
Standout Features & Performance
1. Long-Horizon Autonomy [4]
Fable 5 can plan across distinct operational stages, delegate sub-tasks to smaller parallel sub-agents, and dynamically check its own work over the course of several days without user intervention. In a trial with Stripe, the model migrated a complex 50 million-line Ruby codebase in one day—a process that typically takes humans months. [5, 6, 7, 8]
2. Specialized Domain Benchmarks
According to official reports from Anthropic and early platform testers, Fable 5 sets a new state-of-the-art across complex benchmarks: [9]
- SWE-Bench Pro (Coding): Scores 80.3%, decisively beating Claude Opus 4.8 (69.2%) and GPT-5.5 (58.6%).
- OfficeQA Pro (Enterprise QA): Achieves 57.9% correctness on multi-modal document search and code execution pipeline tests.
- Core Analytics (Finance): Breaks the 90% threshold on long-running analytical filing tasks. [6, 7, 10]
3. Advanced Multimodal Vision [4]
The model interprets dense technical blueprints, complex charts, and financial spreadsheets with higher accuracy than prior generations. It uses standard command-line tools to crop, unblur, or flip incoming source images autonomously. In a showcase of visual capability, Fable 5 successfully played and completed Pokémon FireRed relying purely on raw visual context. [3, 4, 9]
The “Silent Fallback” Safety System
Because Fable 5 shares the raw architecture of Mythos 5—which possesses extreme cybersecurity exploit and biology capabilities—Anthropic built an automated domain-specific fallback system. [2, 7, 11]
- How it works: Automated safety classifiers silently evaluate every prompt session.
- Rerouting: Queries flagging high-risk vectors (like cyberattack routes or biological weapon synthesis) are automatically routed down to Claude Opus 4.8.
- Triggers: This automated redirection happens in fewer than 5% of total sessions on average.
- Billing: Users are not charged the premium Fable 5 token rate for any prompt that triggers an Opus 4.8 fallback. [3, 7]
Availability and Access Pathways
Developers and enterprises can immediately leverage
claude-fable-5 across major cloud ecosystems, including the Anthropic Claude Platform, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. [2]For standard consumer subscription tiers (Pro, Max, Team, and Enterprise), Fable 5 is included at no extra cost during an initial rollout window extending through June 22, 2026. On June 23, it will be removed from standard tiers and shift to credit-metered access while Anthropic monitors compute capacity. [9]
Introducing Claude Fable 5