Ultrathink in Claude Code: What It Actually Does in 2026

Ultrathink is a prompt keyword that, in 2025 versions of Claude Code, triggered the maximum extended thinking budget of 31,999 tokens. As of July 2026 the keyword no longer appears in the official Claude Code documentation or changelog: thinking is now adaptive on current Claude models, and reasoning depth is controlled with the /effort command (low through xhigh), the --thinking CLI flag, and the MAX_THINKING_TOKENS environment variable.

What Is Ultrathink?

Ultrathink is the strongest of the "magic words" that early Claude Code versions scanned for in your prompt to decide how much extended thinking budget to allocate before acting. Extended thinking is the model's internal reasoning phase: gray scratchpad text where Claude works through a problem before writing the visible answer or making tool calls.

The keywords became famous through Anthropic's April 2025 engineering post, Claude Code: Best practices for agentic coding, which recommended asking Claude to plan first and noted that "think" < "think hard" < "think harder" < "ultrathink" each mapped to increasing thinking budgets. The word stuck because it worked and because it was fun to type. What most people asking about ultrathink in 2026 actually want is the current, supported way to make Claude Code reason harder, which is what the rest of this guide covers.

The 2025 Thinking Keywords and Their Budgets

The original keyword-to-budget mapping was hardcoded, and the exact numbers were reverse-engineered from Claude Code's minified JavaScript by Simon Willison on April 19, 2025. His analysis found three tiers:

Two details matter for interpreting this history. First, the mapping was a Claude Code implementation detail, not a model feature: the CLI translated your keyword into the API's budget_tokens parameter. Second, the budget was a ceiling, not a command; the model could stop thinking early. Both details explain why the mechanism aged out once the underlying API changed.

What Ultrathink Does in Claude Code Today

As of July 2026, the string "ultrathink" does not appear in the official Claude Code documentation at code.claude.com or in the project's public changelog, and the fixed 31,999-token budget it once set is gone. The reason is an API-level shift: on current Claude models, manual thinking budgets (budget_tokens) are deprecated or rejected outright in favor of adaptive thinking, where the model itself decides when and how much to think, scaled by an effort setting.

In practice that means thinking is on by default in Claude Code sessions with supported models, without any keyword. Typing "ultrathink" in a prompt today is harmless, and because instruction-following models respond to emphasis, phrases like "think hard about the edge cases" still legitimately nudge the model to reason more. But there is no documented guarantee behind the keyword anymore. Some community writeups report that recent builds map "ultrathink" to a higher effort level for the next turn; we could not verify that in the official changelog, so treat it as unconfirmed and use the documented controls below when the behavior matters.

How to Control Thinking in 2026

Claude Code now exposes reasoning depth through effort levels and thinking toggles rather than prompt keywords. These are the documented controls as of July 2026:

At the API level, the same spectrum is exposed as output_config.effort with values low, medium, high, xhigh, and max, documented in the effort parameter reference. Effort controls more than thinking depth: lower levels also produce fewer, more consolidated tool calls and terser output. You can watch the current state live, since the Claude Code statusline JSON exposes both effort.level and thinking.enabled.

Token and Cost Implications

Thinking tokens are billed as output tokens at the model's standard output rate, and per Anthropic's extended thinking documentation you are charged for the full thinking process even when the interface shows only a summary or nothing at all. Hiding thinking reduces latency and clutter, not cost.

The arithmetic is why "always ultrathink" was never a free lunch. At Claude Opus 4.8's output rate of $25 per million tokens (as of mid-2026), a prompt that actually consumed the old 31,999-token ultrathink ceiling could add roughly $0.80 in thinking alone, before a single line of visible output. Multiply that across a long agentic session and effort settings become a real budget lever. On subscription plans the cost shows up differently, as faster consumption of your 5-hour and weekly usage limits. The practical guidance: run at the default effort for routine work, reach for xhigh on genuinely hard problems, and drop to low or medium for mechanical edits. Our Claude model comparison covers which model tiers support which effort levels.

When Deeper Thinking Actually Helps

Extended thinking pays off on tasks with real search depth: architectural decisions, debugging with several plausible causes, multi-step migrations, and planning before large edits. Anthropic's own best-practices guidance has been consistent since 2025 on the pattern that works: ask Claude to research and plan first, with thinking emphasized, then execute the plan as a separate step. That workflow survives intact in 2026; only the trigger changed from a keyword to an effort setting.

Deeper thinking is wasted on lookups, renames, boilerplate, and anything where the first answer is almost always right. It can even hurt: more deliberation means more latency, and on trivial tasks high effort sometimes produces overthought, overbuilt solutions. If you want structured multi-step reasoning as an inspectable artifact rather than a hidden scratchpad, a tool like the Sequential Thinking MCP server is a complementary approach, and our Claude Code best practices guide covers when to plan versus just act.

Limitations

The honest caveat about ultrathink in 2026 is that keyword-triggered behavior is undocumented behavior. Claude Code has changed thinking mechanics several times (fixed budgets, then defaults, then adaptive thinking and effort), and anything you learned from a 2025 blog post, including the 4,000/10,000/31,999 numbers above, describes a version you are probably not running. Verify against the current docs before building habits or scripts around a keyword.

Also remember that thinking quality is bounded by context quality. No effort level rescues a prompt that is missing the constraint that matters, and a bloated context window degrades reasoning faster than a low effort setting does. Fix your CLAUDE.md and your prompt first; turn the effort dial second.

Related Guides

• Claude Code Guide [2026]: CLI, Desktop, Web, and Real-World Fit
• Claude Code Best Practices in 2026: What Actually Holds Up
• Best Claude Models in 2026: Sonnet vs Opus vs Haiku Compared
• Claude Code MCP: How to Add and Manage MCP Servers