Ethan Mollick reviews Claude Fable 5

Fable's ability to act as a self-directing multi-agent studio signals the death of prompt engineering and the rise of result-based oversight, but its astronomical token cost and over-aggressive safety filters will throttle early adoption.

• –**Autonomous Sub-agent Orchestration:** The model's most defining feature is its ability to spin up, manage, and verify results from a network of cheaper sub-agents (like Sonnet) to handle massive research and testing loads.
• –**Black-Box Shift:** As AI assumes control over thousands of micro-decisions during multi-hour execution windows, humans lose granular steerability and transition into passive evaluators.
• –**Aggressive Security Fallbacks:** Fable's guardrails are highly sensitive, causing frequent silent fallbacks to the older Claude Opus 4.8 for benign prompts, creating user friction.
• –**High Resource Consumption:** The model operates at double the cost of Opus and consumes immense token volumes, making production efficiency a key future bottleneck.