Why Claude Fable can finish in minutes without producing the longest answer

Claude Fable 5.1 can feel surprisingly fast for Anthropic’s top generally available model because “maximum intelligence” is not a promise to occupy a fixed number of minutes or produce a book-length answer. Anthropic’s effort control is a behavioural instruction that lets the model spend more tokens and explore more deeply when the task requires it. Adaptive reasoning can still decide that a task is complete after a short path.[2][3]

The product default also matters. Anthropic says Fable 5.1 defaults to High effort in Claude Code but Medium in Claude.ai and Cowork. A user who opens the consumer product and assumes it is running the same default as the coding agent may be comparing two different operating points. Lower effort can still perform strongly: Anthropic says Low or Medium Fable 5.1 can match or exceed Fable 5 on some results at lower cost.[1]

Reasoning depth and answer length are not the same variable

A model can reason internally, choose a compact solution and return a short response. It can also produce a long answer with shallow repetition. Anthropic’s effort documentation says the setting influences thinking, response text and tool use, while max_tokens is a hard ceiling for total thinking plus answer output in the API. Setting Max effort without providing a sufficiently large output allowance can therefore constrain the run; providing a large allowance still does not require the model to consume it.[3]

Prompt design strongly affects extensiveness. “Audit this pack and return the import file” invites a bounded artifact. “Research every relevant source, compare alternatives, explain every decision and deliver a full validation record” defines a larger job. If the user asks only for the end product, a capable model may compress intermediate analysis instead of turning it into public prose.

More effort can have diminishing returns

Anthropic’s own performance-and-cost analysis shows why the top setting need not be the automatic choice. On one reported Fable 5.1 evaluation, moving from Low to Max increased the score substantially, but the last step to Max added about half a point while raising cost by 46%, with the gain described as inside run-to-run noise. That is one benchmark rather than a universal law, but it demonstrates that additional effort can buy little on a task that is already near its useful ceiling.[5]

A quick response can reflect the selected effort, concise output, fewer tool calls or a narrower interpretation of the task. To compare products fairly, give them the same files, tool access and acceptance criteria, then record the correction time as well as the initial run. The published defaults identify one variable to control; they do not establish which product will finish a particular job best.[1]

The practical comparison is not “which one thought longest?” It is which system produced the accepted artifact with the least total time, including correction. A five-minute answer that needs an hour of repair can lose to a twenty-minute answer that imports cleanly. Conversely, a fifty-minute run that repeats itself without improving the result is not automatically more intelligent. Speed, extensiveness and accepted quality should be recorded separately.

Report an error or suggest a correction