Claude Code Extended Thinking: How Ultrathink Mode Actually Works

Extended thinking is a mode where Claude Code reasons through a problem in a visible thinking pass before it writes its actual response, instead of going straight from your prompt to an answer. That extra pass costs tokens and time, and it buys you a model that has worked through tradeoffs, checked its own logic, and considered edge cases before committing to a plan. For a quick file rename, you do not need that. For a decision that will shape the next month of a codebase, you do.

You trigger extended thinking by putting one of four keywords anywhere in your prompt, and each one asks for a bigger reasoning budget than the last: think, then think hard, then think harder, then ultrathink at the top. Ultrathink was re-introduced in Claude Code v2.1.68 and gives the agent roughly 32,000 tokens to reason through the problem before it responds. The keyword is a one-time nudge for that specific message - it does not change your session's effort setting and does not carry over to your next prompt.

The extra token budget goes toward the agent working through a problem the way a careful engineer would on a whiteboard before touching the keyboard: naming the options, checking each one against the actual constraints, and catching a flawed assumption before it turns into code. That is worth the wait on decisions where being wrong is expensive to unwind - choosing between two database schema approaches, tracking down a bug that only reproduces intermittently, or reviewing a security-sensitive piece of code for what an attacker could actually do with it.

Do not reach for ultrathink to fix a typo or rename a variable. Extended thinking adds real wait time for a bigger reasoning budget, and on a task with one obvious right answer, that budget has nothing useful to do with the extra tokens. Save it for the decisions that actually have more than one reasonable path, and let routine work move at normal speed.

The Effort Ladder is the checklist I run before typing a thinking keyword at all: match the level to how expensive it would be to get the decision wrong, not to how important the project feels in general.

The newer Claude models are starting to handle some of this automatically. Adaptive thinking lets the model decide on its own how deeply to reason based on how complex your request actually is, rather than waiting for you to type a keyword. Anthropic recommends adaptive thinking as the standard setting for the newest Opus and Sonnet models. That does not make the four keywords useless - they are still the explicit override when you want to force a deeper pass on a specific call - but expect the model to lean on adaptive reasoning by default more often as this keeps evolving. Check your version's release notes before assuming last month's behavior still applies.

I had a project where the obvious data model was a single table with a status column, and it would have worked fine for the first version. Before committing to it, I asked Claude Code to think harder about whether that shape would hold up once a second workflow needed to read the same records in a different order. The extra reasoning pass surfaced a real problem - two features I had planned would end up fighting over the same status field - and it proposed a small split that avoided the conflict entirely. That is not a dramatic story. It is exactly the kind of quiet, avoided rework that extended thinking is actually for: the mistake you never had to notice because it never got built.