Claude Ultrathink: What It Does in Claude Code Now
Claude ultrathink explained: what the keyword does in Claude Code today, how it relates to effort levels and plan mode, and when the extra tokens pay off.
Josh Pocock6 min read
AI Coding
Ultrathink
Stride AI Academy
Claude ultrathink is a keyword you type into a Claude Code prompt to ask for deeper reasoning on that one turn. Claude Code spots the word and adds an instruction to think harder, without changing your session's effort setting. In the terminal, VS Code and the desktop app's Code tab, it also raises that turn to the xhigh effort level. Other phrases such as "think hard" are no longer keywords.
This guide covers what the keyword does now, what changed from older advice, how it fits with effort levels, plan mode and cost, and when it's worth using. Everything here was checked against Claude Code's model configuration docs, the costs docs and the Claude API's extended thinking docs on November 18, 2026.
What ultrathink does today
Put the word anywhere in your prompt:
ultrathink: why does the checkout total drift by a cent on some orders?According to the docs, two things can happen on that turn:
- An in-context instruction. Claude Code recognises the keyword and adds an instruction asking for deeper reasoning. This applies wherever the keyword is recognised.
- A jump to
xhigheffort. When you type the word in the terminal prompt, VS Code's chat box or a local session in the desktop app's Code tab, Claude Code also runs that turn atxhigh. In the terminal and VS Code, the word shows in rainbow colours when it can raise the turn.
Your session's effort level stays where it was. The next prompt without the keyword goes back to normal.
| Where you type it | Raises the turn to xhigh from |
|---|---|
| Terminal prompt | Claude Code v2.1.311 |
| VS Code chat box | Claude Code v2.1.316 |
| Desktop app Code tab (local session) | Claude Code v2.1.316 |
What changed from older advice
Older guides describe "think", "think hard" and "ultrathink" as rising tiers of thinking budget. That isn't how the current docs describe it. Today, Claude Code passes "think", "think hard" and "think more" through as ordinary prompt text and doesn't treat them as keywords. Only ultrathink is recognised.
The budget idea itself has moved on too. The Claude API's extended thinking page now calls fixed budget_tokens budgets a legacy feature, and models numbered 4.7 and later reject them. Newer models use adaptive reasoning instead, where Claude decides how much to think on each step and the effort level sets how deep it goes.
If the word doesn't turn rainbow
The docs list three fixes when the keyword isn't raising the turn:
- Retype it in plain text. Keep it outside quotes, brackets, a file path or a command, in a prompt that doesn't start with
/. - Check your effort cap. Unset
CLAUDE_CODE_EFFORT_LEVELif it's set belowxhigh, raise your ownmaxEffortLeveltoxhighor above, or ask an admin to raise your organization's limit. - Check the model. Run
/modeland pick Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.2 or Fable 5.1.
On Amazon Bedrock, Google Cloud's Agent Platform or a Claude apps gateway, the word can't raise a turn. Set the effort level directly instead.
Ultrathink vs effort levels
Effort is the main dial for how much Claude reasons. Ultrathink is a one-turn bump. The docs list these levels:
| Level | What the docs say it's for |
|---|---|
low |
Quick exchanges you review closely, or small changes |
medium |
The default on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.2 |
high |
Work where verification matters, such as a bug fix in an existing codebase |
xhigh |
Deeper reasoning at higher token spend; the default on Opus 4.7 |
max |
Hard problems you want Claude to work through alone; can overthink |
Change the session level with /effort:
/effort high
/effort status
/effort autoOr set it at launch or in your shell:
claude --effort high
export CLAUDE_CODE_EFFORT_LEVEL=highYou can also save a default with the effortLevel key in settings. max isn't accepted there and applies to the current session only, unless you set it through the environment variable.
A simple rule: if you want deeper reasoning for one hard question, type ultrathink. If a whole session needs it, raise the effort level.
Thinking toggles and MAX_THINKING_TOKENS
A few related controls still exist, but they behave differently by model:
- Alt+T (Windows and Linux) or Option+T (macOS) toggles thinking for the current session.
/configsets the default, saved asalwaysThinkingEnabled. - Ctrl+O toggles verbose mode so you can read Claude's thinking as grey italic text. It's collapsed by default.
MAX_THINKING_TOKENSonly sets a budget on models with a fixed thinking budget. Adaptive models ignore nonzero values. Setting it to0disables thinking where that's allowed.
You can't turn thinking off on Opus 5.5, Sonnet 5.5, Haiku 5.5 or the Fable models, so on those, effort and ultrathink are the controls that matter.
How it relates to plan mode
Plan mode controls what Claude is allowed to do: it reads and proposes a plan, and makes no edits until you approve. Ultrathink controls how hard Claude reasons on a turn. They solve different problems, so they combine well. Switch to plan mode with Shift+Tab, then ask for the plan with the keyword:
ultrathink and plan how to move our auth from sessions to JWTs without logging anyone outThe costs docs recommend plan mode for complex tasks because a wrong first direction leads to expensive rework. Spending extra reasoning on the plan, rather than on every edit afterwards, is usually the cheaper place to put it.
What it costs
Thinking isn't free. The docs say thinking tokens are billed as output tokens, and you're charged for all of them, even when the thinking is collapsed or redacted. A turn at xhigh will usually spend more than the same turn at medium.
On a Pro, Max or Ultra plan, that comes out of your usage limits. On the API, it's on your bill. For the full picture of plans and API pricing, see how much Claude Code costs, or estimate a month with the cost calculator. Long sessions cost more on every turn too, so pair it with good context management.
When ultrathink is worth the tokens
Use it for turns where a better answer saves real time:
- A bug you've already tried to fix once and it came back
- Planning a change that touches many files or a database schema
- Choosing between two designs where the trade-offs aren't obvious
- Reviewing a tricky diff for edge cases before you merge
Skip it for renames, formatting, small copy edits and anything you'll check line by line anyway. Those are low or medium work.
To stop the keyword working at all, including its in-context instruction, set this in your shell or a settings file (Claude Code v2.1.316 or later):
export CLAUDE_CODE_DISABLE_ULTRATHINK=1Learn to build with Claude Code
The Stride AI Coding course teaches Claude Code and Codex for building client tools, written for non-programmers, with weekly live calls to review your builds. Many of the free guides cover Claude Code setups, and the Claude Code community is where members compare theirs.
FAQ
What does ultrathink do in Claude Code?
It asks Claude for deeper reasoning on the turn where you type it, by adding an in-context instruction. In the terminal, VS Code and the desktop app's Code tab, it also raises that turn to xhigh effort. Your session's effort setting doesn't change.
Does "think hard" still work in Claude Code?
Not as a keyword. The current docs say Claude Code passes "think", "think hard" and "think more" through as ordinary prompt text. Only ultrathink is recognised.
Is ultrathink the same as max effort?
No. When it raises a turn, it raises it to xhigh, and only for that turn. max is a separate effort level you set with /effort max, and it lasts for the session.
Why isn't the ultrathink word showing in rainbow colours?
Check that it's outside quotes, brackets, paths or commands, that CLAUDE_CODE_EFFORT_LEVEL isn't capping you below xhigh, and that you're on a supported model such as Opus 5.5 or Sonnet 5.5. On Bedrock, Google Cloud's Agent Platform or a Claude apps gateway, set the effort level instead.
Does ultrathink cost more?
Usually, yes. Thinking tokens are billed as output tokens, and a deeper reasoning turn tends to produce more of them. Use it on hard turns, not routine ones.
How do I turn ultrathink off?
Set CLAUDE_CODE_DISABLE_ULTRATHINK=1 in your shell or in a settings file. It requires Claude Code v2.1.316 or later and disables both the keyword and its in-context instruction.