EVERYTHING
Anthropic has introduced Claude Fable 5, a Mythos-class model built explicitly for handling long-running autonomous operations. According to Futurumgroup, this release marks a significant advancement in the capabilities of AI agents, moving beyond simple prompt execution toward sustained, multi-hour work blocks.
Sustained Autonomy Redefines Agent Capabilities
The core innovation of Fable 5 lies in its duration and scope. It is designed to execute asynchronous tasks over extended periods inside an agent harness, allowing it to plan, check progress, and refine code iteratively. The model boasts a massive 1 million token context window and can generate up to 128,000 output tokens per request, enabling it to reconstruct application code from screenshots and complete long-horizon development tasks that previous models could not manage.
This capability forces engineering teams to rethink their delegation strategy. The practical question shifts from how well a model performs on one specific prompt to how long it can carry out an entire development task with human oversight. When a model achieves this level of sustained autonomy, the natural unit of work moves away from a single pull request and toward a bounded work block that the agent owns entirely.
Pipeline Design and Governance Challenges
The deployment of Fable 5 across various platforms—including the Claude API, AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry—introduces new considerations for pipeline security and governance. The model is priced at $10 per million input tokens and $50 per million output tokens.
Crucially, Fable 5 ships with integrated safety classifiers that allow it to decline certain requests. This refusal mechanism is designed into the API workflow: when a request is declined, the Messages API returns a successful response naming the classifier responsible. Teams can then configure a fallback system, such as retrying the request on another model like Opus 4.8.
- Safety and Fallback: Harmful prompts in sensitive areas, including cybersecurity, biology, chemistry, and health, are routed to more restricted models.
- Classifier Efficacy: Anthropic reports that these safeguards trigger in fewer than 5% of sessions.
- Model Variants: A version of the model without safety classifiers (Claude Mythos 5) remains limited to approved customers through Project Glasswing.
This integrated refusal-and-fallback boundary places particular importance on software security work, as the classifier that blocks offensive cyber activity also intersects with defensive AppSec and supply chain tasks engineers manage. Ultimately, Fable 5 compels organizations to redesign their workflows to account for model switching mid-task and to track model provenance throughout a complex, unattended run.