Extended Thinking & Effort Levels

what extended thinking actually is

Extended thinking is the reasoning Claude produces before it answers. Claude Code collapses it by default; press Ctrl+O to open the transcript viewer and see it as gray italic text. Hiding it doesn't change what it costs — every thinking token generated is billed whether or not you ever expand the view.

the "think harder" myth

Advice that typing think, think hard, or

think
                harder

into a prompt makes Claude reason more deeply circulates constantly — and in current Claude Code, none of it does anything. Those phrases pass straight through as ordinary prompt text with no special effect on how much the model reasons.

The one keyword actually recognized is ultrathink. Put it anywhere in a message and Claude Code adds an in-context instruction requesting deeper reasoning for that single turn — without touching your session's effort setting.

effort levels: the real control

Effort level is what actually governs how much a model reasons, on any model that supports it. Run /effort with no arguments for an interactive slider, /effort <level> to set one directly, or /effort auto to reset to the model's default.

Level	When to use it
`low`	short, scoped, latency-sensitive tasks
`medium`	cost-sensitive work that can trade off some intelligence
`high`	the default on nearly every model — balances cost and intelligence
`xhigh`	deeper reasoning at higher token spend
`max`	the deepest reasoning available; prone to overthinking simple tasks, session-only unless set via environment variable

The scale is calibrated per model, so the same level name doesn't represent the same underlying depth on every model — and requesting a level a given model doesn't support falls back to the highest one it does.

adaptive reasoning

On current models, Claude decides per step whether thinking is worth it at the active effort level, rather than spending a fixed thinking budget on every single step regardless of need. Effort level shapes how often deep reasoning kicks in, not a fixed token allowance. If you want more or less thinking than a level is producing, say so directly in the prompt or in CLAUDE.md — Claude responds to that guidance within its current effort setting.

toggling and cost

Control	How to set it
toggle for the current session	`Option+T` (macOS) / `Alt+T` (Windows/Linux)
set the global default	`/config` → thinking mode
view thinking output	`Ctrl+O` to open the transcript viewer

Toggling the display changes what you see, not what you pay — thinking tokens are billed the same whether the reasoning is expanded, collapsed, or never opened at all. On models that don't allow thinking to be disabled, the session toggle simply has no effect.

reference

code.claude.com — model configuration, effort levels & extended thinking
platform.claude.com — effort levels