Ultrathink Is Back in Claude Code: What Changed and How to Use It
Claude Code v2.1.68 restores the ultrathink keyword after a controversial January removal. Here's what it does now, how the new effort level system works, and why it matters.
In this article
What Ultrathink Is
Ultrathink is a keyword you type directly into your Claude Code prompt. When Claude Code detects it, it sets the thinking effort for that turn to high on Opus 4.6 and Sonnet 4.6 — giving the model more room to reason before it responds. You can use it anywhere in a prompt: "ultrathink this refactor" or "help me design this API, ultrathink" both work the same way.
The visual cue is back too. When you type ultrathink, Claude Code highlights the keyword to confirm extended thinking is active. Claude Code creator Boris Cherny noted on Threads: "When you type a keyword like 'think' or 'ultrathink', Claude Code will highlight it to show that extended thinking is enabled. Hit /t to turn it off, so you can include those words in your prompt without triggering extended thinking by accident."
Claude Code only. The ultrathink keyword triggers special behavior exclusively in the Claude Code CLI. Typing it into claude.ai or the API does nothing special — it's treated as regular text.
The History: Deprecation and Backlash
For most of its life, ultrathink worked by unlocking a maximum thinking budget of 31,999 tokens. Typing it into a prompt was the easiest way to get Claude's deepest reasoning. Power users built workflows around it, and the "Opus + Ultrathink + Plan Mode" combination became a standard recommendation for complex tasks.
On January 16, 2026, Anthropic removed it. The reasoning: extended thinking was now enabled by default for all supported models at the maximum budget, so the keyword was redundant. The underlying code that inspected prompts for thinking keywords was removed entirely.
What followed surprised the team. GitHub issue #19098 — "Restore explicit ultrathink keyword as opt-in extended thinking mode" — accumulated significant developer feedback about quality degradation after the change. The complaint wasn't that thinking was disabled; it was that Claude Code's new automatic thinking allocation appeared to sometimes optimize for cost or speed rather than thoroughness. Users paying Opus 4.6 prices reported outputs that felt inconsistent with what they expected from the model.
A separate issue (#18884) flagged the removal of the rainbow text highlighting that appeared when ultrathink was typed — a smaller UX detail, but one that pointed to how much the keyword had become part of the Claude Code ritual for many developers.
Both issues were closed as completed in February 2026, and v2.1.68 shipped with ultrathink restored.
How It Works Now: The Effort Level System
The restored ultrathink operates differently than the original. Rather than directly controlling a token budget, it's now a per-turn shortcut within Claude Code's new effort level system.
Effort levels (low, medium, high) control how aggressively Opus 4.6 and Sonnet 4.6 allocate thinking before responding. You can set your default effort via /model in Claude Code, or with the CLAUDE_CODE_EFFORT_LEVEL environment variable. Opus 4.6 defaults to medium effort for Max and Team plan subscribers.
- Low — Fewer thinking tokens, faster responses, lower cost
- Medium — Balanced behavior, the default for most plan types
- High — Maximum reasoning depth; what ultrathink triggers for that turn
Ultrathink sets effort to high for that single turn without changing your session default. This is the key design change from the original: it's an override, not a global toggle. Use it when a specific prompt needs deep reasoning but you don't want every prompt in your session to run at maximum effort and cost.
Important note from the docs: "Phrases like 'think', 'think hard', and 'think more' are interpreted as regular prompt instructions and don't allocate thinking tokens." Only ultrathink triggers the effort-level override. The lesser thinking keywords no longer have any mechanical effect.
Supported Models
Ultrathink's effort-level behavior applies to Opus 4.6 and Sonnet 4.6. For other models, Claude Code uses a fixed thinking token budget rather than the adaptive reasoning system, so the effort level mechanics don't apply in the same way.
Sonnet 4.6 is now the default model for Pro and Team plans. At roughly one-fifth the cost of Opus and within ~1 percentage point on SWE-bench, it makes ultrathink particularly useful on Sonnet: you can run at medium effort by default for most tasks, then invoke ultrathink for the prompts that actually need it — getting near-Opus reasoning depth without paying Opus prices across the board.
Additional Thinking Controls
Ultrathink is one part of a broader set of thinking controls in Claude Code:
- Option+T / Alt+T — Toggle thinking on or off for the current session (all models)
- /config — Set your global thinking default (saved as alwaysThinkingEnabled in ~/.claude/settings.json)
- CLAUDE_CODE_EFFORT_LEVEL — Environment variable to set default effort (low, medium, high) for Opus 4.6 and Sonnet 4.6
- MAX_THINKING_TOKENS — Limit thinking to a specific token count; set to 0 to disable thinking entirely on any model
- Ctrl+O — Toggle verbose mode to see Claude's internal reasoning displayed as gray italic text
For developers building on the API, adaptive thinking (thinking: {type: "adaptive"}) is the recommended mode for both Opus 4.6 and Sonnet 4.6. The effort parameter maps to the same low/medium/high controls available in Claude Code. MAX_THINKING_TOKENS is ignored on Opus 4.6 and Sonnet 4.6 unless set to 0.
When to Use It
The official documentation calls out these as the scenarios where ultrathink adds the most value: complex architectural decisions, challenging bugs, multi-step implementation planning, and evaluating tradeoffs between different approaches.
The practical test is simple: if you're about to prompt Claude Code to make a decision that would be annoying to undo or revisit, ultrathink first. Planning a refactor across multiple files, designing a new data model, or evaluating which of several approaches to take — these are the cases where spending the extra tokens on reasoning depth pays off. Routine code generation, formatting changes, or documentation don't need it.
Boris Cherny's personal recommendation: use high effort for everything by default. Run /model in Claude Code and set your effort level to high as a session default if your workflow involves complex tasks throughout. Ultrathink is then useful as a reminder or emphasis signal rather than a major mode switch.
Quick Takeaway
Ultrathink is back in Claude Code v2.1.68 after being removed in January 2026. It now works as a per-turn override that sets effort to high within the new effort level system, supported on Opus 4.6 and Sonnet 4.6. The keyword triggers visual highlighting in the Claude Code prompt to confirm it's active. Community feedback on GitHub — specifically issue #19098 documenting quality degradation after the original removal — directly drove its restoration.
The clearest use pattern: leave your session at medium effort for routine work, invoke ultrathink for the specific prompts where you need maximum reasoning depth. On Sonnet 4.6, this gives you a practical way to access near-Opus reasoning on demand without committing to Opus pricing for every turn.
Read the extended thinking documentation on code.claude.com
Get the Claude playbook in your inbox.
One weekly email for Claude and Claude Code users. Real workflows, no hype. Subscribe and we send you The Claude Power-User Cheatsheet.
— ¶ —
Luke Thompson
Luke Thompson is the founder of The Operations Guide, LLC and editor of The Claude Insider. Based in Jonesborough, Tennessee, he has spent years building AI-augmented business systems and automation workflows for operators and teams. He began working with large language models in production well before the current wave of consumer AI tools, integrating them into client workflows, content pipelines, and operational infrastructure. At The Claude Insider, he writes about Claude with the specificity of someone who uses it daily as a professional tool — not as a reviewer or commentator, but as a builder. His coverage focuses on what actually works: prompt patterns, API integration strategies, agentic workflows, and the real-world tradeoffs that practitioners face. He is not affiliated with Anthropic, PBC.
Articles are researched and drafted with AI assistance, reviewed and edited by Luke Thompson.
Know where AI can pay off in your company.
Take the free two-minute AI Readiness Assessment. See your score, the two gaps holding you back, and the next move worth making.
