Earlier Claude models let you turn extended thinking on or off and set a token budget for it. Claude Fable 5.1 removes all of that. It thinks before every response, the depth is controlled by effort, and the raw chain of thought is never shown.

Why you can't turn it off

The thinking is part of how the model gets its results. Anthropic made it structural rather than optional. On the API, sending any thinking configuration other than "adaptive" — including "disabled" — returns an error.

For you, that means the question "should I enable thinking for this?" no longer exists. The question is "how much effort does this task deserve?", which is a better question.

What you see instead of the reasoning

By default, nothing. The model thinks, then answers, and the thinking is invisible. If your tool shows a "thinking" section, it's either empty or a summary — never the actual reasoning.

You can ask for a summarised view, which returns a readable digest of what it considered. That's useful when a long turn goes quiet and you want to know what it's doing. In Claude Code and similar agent tools, there's also a progress-update mode that surfaces its between-tool-call notes as short status lines.

If the tool you're using shows a long silence before output, that's the default omitted mode. It's not stuck.

What this means for how you work

Stop trying to see its reasoning to check it. Check the output instead. Ask it to explain a decision, ask for the plan before the code, ask it to audit its own claims against evidence — all of those work. Reading the raw thinking never did, and now you can't.

Longer pauses are normal. Especially at high effort. The pause is the work.

Don't paste thinking back in. If you're building your own tool on the API, pass the model's thinking blocks back unchanged in multi-turn conversations. Don't edit them, don't strip them, don't reorder earlier turns — Fable 5.1 checks that the history hasn't been tampered with and rejects edited conversations.

The one behaviour to know about

Because it always thinks, it can occasionally over-plan on an ambiguous task — weighing options at length when you just wanted it to pick one. A short nudge fixes it:

act-dont-deliberate.txt
When you have enough information to act, act. Don't re-derive facts
already established, re-litigate a decision I've made, or narrate
options you won't pursue. If you're weighing a choice, give me a
recommendation, not a survey.

If you find yourself wanting to watch it think, that's usually a sign you don't trust the output yet. Build the trust with tests and a review pass, not by reading reasoning you can't see anyway.