Anthropic's Claude Fable 5 Keeps Reasoning Always On, Lets Developers Adjust Depth

Anthropic has introduced Claude Fable 5, a hybrid reasoning model that keeps its internal thinking process permanently active for every request. Unlike earlier large language models that processed all queries at a uniform compute level, Fable 5 allocates reasoning effort based on a developer-controlled setting. Developers can set the effort parameter to one of five levels — low, medium, high, xhigh, or max — through an output_config block in the updated API, replacing the older token-budget toggle. The always-on thinking phase generates internal reasoning tokens before producing a final response, which Anthropic says improves accuracy for mathematics, coding, and multi-step logic tasks. Attempting to disable thinking via older API parameters returns an HTTP 400 error, as those controls have been removed from the schema.