- Home
- Learn
- Deep Dives
- Claude Ultrathink, Ultraplan and Effort Levels Explained
Claude Ultrathink, Ultraplan and Effort Levels Explained
What ultrathink does now, why Ultraplan is gone and how effort levels and adaptive thinking set Claude's reasoning and cost. Checked 13-11-2026.
Published 13-11-2026. Last verified on 13-11-2026 against the Anthropic API and Claude Code documentation, reviewed monthly because these controls change with almost every model release.
Quick answer: what controls how hard Claude thinks today?
This guide is for developers and team leads who want hard problems to get enough reasoning from Claude while routine work does not burn tokens, and who keep running into advice about ultrathink and ultraplan that no longer matches what they see. The short version, as of 13-11-2026: effort levels are the main control. You set low, medium, high, xhigh or max with /effort in Claude Code or output_config.effort in the API, and on current models Claude decides per request how much to think within that level. Ultrathink still exists in Claude Code as a one-turn boost, not as the fixed token budget older guides describe. Ultraplan has been removed. Extended thinking with a fixed budget_tokens is a legacy feature that the newest models reject. Thinking is billed as output tokens, so every step up in effort can show up on the invoice.
How we checked this
Our method in three lines: every statement about a control is quoted from the Anthropic API docs or the Claude Code docs as we read them on 13-11-2026; prices come from the vendor's pricing page and any cost on this page is our calculation on those prices; and what we did not test is labelled as such. The gate for this page: a control earns a recommendation only if the official docs describe a concrete effect we can point to. Our original plan was to run one hard task at every setting and compare tokens, time and correctness. This run had no model access, so that comparison is not tested by us, and we say so rather than guess.
The gate removed two things that still circulate. The 2025 keyword ladder of think, think hard and think harder is out as a current control, because the Claude Code model configuration docs say those words do not act as triggers. And Ultraplan is out as a recommendation, because Anthropic's own page now says it is no longer available.
What does ultrathink do in Claude Code now?
In current Claude Code, typing ultrathink in a prompt raises the effort for that one turn without changing your session's effort setting. The Claude Code model configuration docs, as we read them on 13-11-2026, describe it as a per-turn boost, and in a terminal on version 2.1.311 or later it can take that turn to xhigh. Words such as think hard or think more are not triggers.
That is a change from where the keyword came from. Community analyses of Claude Code in April 2025 reported that think, think hard and ultrathink mapped to fixed thinking budgets of roughly 4,000, 10,000 and 31,999 tokens. Two of the three pages that rank for this topic still present those numbers as current. They describe a mechanism that the newer models replaced with effort levels and adaptive thinking.
Our advice: use ultrathink for the single hard question inside an otherwise routine session, such as "why does this race condition only happen under load", and leave the session effort where it is. If you find yourself typing it in every prompt, raise the session effort with /effort instead; that is what the setting is for. Clear prompts still matter more than the keyword. Our guide on writing better prompts for Claude Code covers the habits that make the extra thinking worth paying for.
The exact version number and the xhigh mapping come from a part of the docs page we read through an automated summary, so check them against your version before you rely on them.
What happened to Ultraplan?
Anthropic removed it, and we would stop looking for it. The Ultraplan page in the Claude Code docs now opens with the heading "Ultraplan is no longer available" and says the research preview has been removed: the /ultraplan command, the ultraplan keyword trigger and the option to refine a plan with Ultraplan on Claude Code on the web are gone. The page does not give a removal date.
The replacements Anthropic names on that page are plan mode in your local session, where Claude researches the codebase and presents a plan for your approval before changing anything, and Claude Code on the web for running sessions in the cloud and reviewing changes in the browser. Plan mode starts with claude --permission-mode plan or with Shift+Tab during a session. Our Claude Code plan mode guide walks through it.
Plan mode plus a higher effort level gets you most of what people used Ultraplan for. Two of our skills make the planning step stricter. Plan First, Code Second is free and has the agent write a plan, then failing tests, then the smallest change that passes, with approval gates. Write the Spec Before Any Code, $12, produces a spec, plan and task list and has a second model check the result against the spec. For teams that keep specs as documents, the technical documentation collection has related skills.
What is extended thinking, and is it still used?
Extended thinking is the original API mechanism: you set thinking.type to enabled with a budget_tokens value, and Claude reasons up to that budget before answering. Anthropic now calls it a legacy feature. The API docs we read on 13-11-2026 say that Claude 4.7 and later models do not support it and reject such requests with a 400 error. On Opus 4.6 and Sonnet 4.6 budget_tokens is deprecated, and on Opus 4.5, Haiku 4.5 and older 4-series models it still works, with a minimum budget of 1,024 tokens that must stay below max_tokens.
So in our reading "extended thinking" now means two things. In the API it is the old fixed-budget mode. In Claude Code the term is still used for thinking in general: the docs describe Option+T on macOS or Alt+T on Windows and Linux as the toggle, with no effect on Opus 5.5, Sonnet 5.5, Haiku 5.5 or the Fable models, which always think.
Where we still see it matter: if your team runs scripts or agents against Haiku 4.5 or another older model through the API, budget_tokens is your control. For anything on current models, migrate the code to adaptive thinking and effort before the old parameter breaks a pipeline.
What is adaptive thinking?
Adaptive thinking lets the model decide per request whether to think and how much. The Anthropic docs on thinking steering and cost put it as: the model evaluates each request and determines whether to think and how much. You switch it on with thinking type adaptive and steer the depth with the effort level.
Support depends on the model, and we would check this list before every migration, according to the same docs. On Opus 5.5 adaptive thinking is always on and cannot be turned off; a request with thinking disabled returns a 400 error. Haiku 5.5 can run with thinking disabled at high effort or lower. Opus 4.5, Sonnet 4.5, Haiku 4.5 and older models return a 400 error for adaptive.
Our take: adaptive thinking is the right default for almost all work, because most requests in a coding session are small and do not deserve a long reasoning pass. The cost saving comes from the easy requests where the model thinks briefly or not at all. If you want to see how much a model thought, the API returns the count in usage.output_tokens_details.thinking_tokens.
What are effort levels, and which should you use?
Effort sets how much work Claude puts into the whole response, including how often and how deeply it thinks in adaptive mode. The levels are low, medium, high, xhigh and max, and the API effort documentation describes low as the most efficient with significant token savings, xhigh as extended capability for long-horizon work and max as maximum capability with no constraints on token spending.
Defaults differ per model, which is why we never assume one. From the API docs on 13-11-2026:
| Model | Levels | API default |
|---|---|---|
| Fable 5.1, Fable 5 | low to max, incl. xhigh | high |
| Opus 5.5 | low to max, incl. xhigh | medium |
| Sonnet 5.5 | low to max, incl. xhigh | high |
| Haiku 5.5 | low to max, incl. xhigh | medium |
| Opus 5, Opus 4.8, Opus 4.7 | low to max, incl. xhigh | high |
| Opus 4.6, Sonnet 4.6 | low, medium, high, max (no xhigh) | high |
In Claude Code you set it with /effort, which opens a slider, or /effort high to set it directly; the CLAUDE_CODE_EFFORT_LEVEL variable and the --effort flag do the same for scripts. In our reading of the Claude Code docs, some Claude Code defaults differ from the API defaults, for example medium for Sonnet 5.5, so check the slider rather than assume.
Our picks for a working team, in order of how often we would use them:
- Medium for daily coding and edits. It sits first because most of a working day is small changes, and medium keeps those fast and cheap on the models that default to it.
- High for debugging and design decisions. Second because these are the tasks where a wrong first answer costs more than the extra tokens.
- xhigh for long, multi-step work such as a migration across many files. Third because it pays off only when the task is long enough to need it.
- Low for bulk and mechanical work like renaming or formatting. Fourth because the saving is real but the tasks are rarer than people think.
- Max almost never. Last because the docs describe it as having no constraints on token spending; we would only use it on a single problem that has beaten high and xhigh.
To choose the model as well as the level, we would look at benchmark data first. Your agent can pull model comparisons and cost per task from Artificial Analysis through Agensi Tools, from $0.013 per call for model details and from $0.13 per call for cost per task on 13-11-2026, or check the LMArena leaderboards from $0.013 per call, without an account or API key at either provider. Route Each Task to the Right Model, $12, automates the choice by escalating to a stronger model only when a quality check fails. Our LLM ops collection and the research and analysis collection have more.
We sell skills and pay-per-call tools that your agent uses through one MCP connection, and we name ours in each section. No vendor paid to be listed.
How are thinking tokens billed?
Thinking tokens are billed as output tokens, and you pay for the full thinking process, not only the summary you see. The thinking steering and cost docs say so explicitly, and add that summarised and omitted thinking display cost the same. Earlier thinking blocks that stay in the context are billed again as input on later turns, on models that keep them.
Output is the expensive side, and that is the number we would watch. From the Anthropic pricing page on 13-11-2026, per million tokens:
| Model | Input | Output | 10,000 thinking tokens cost |
|---|---|---|---|
| Opus 5.5 | $4 | $20 | $0.20 |
| Sonnet 5.5 | $2 | $10 | $0.10 |
| Haiku 5.5 | $0.10 (up to 100,000 tokens) | $0.50 (up to 100,000 tokens) | $0.005 |
The last column is our calculation on those vendor prices, not a measurement. It shows why the effort level matters more than the model choice for many teams: twenty hard prompts a day at 10,000 thinking tokens each on Opus 5.5 is $4 a day in thinking alone, before the answer itself. If you use Claude Code on a subscription rather than the API, thinking counts against your usage limits instead of an invoice.
For ways to bring that down, our guide to reducing Claude Code token usage is the practical companion. Cut Agent Token Costs: Caching, Routing and Budget Guards, $5, packages scripts for context hygiene, caching and budget-aware routing. Teams that run Claude in CI can set the level per job; the DevOps and cloud collection has skills for that wiring.
When does more thinking make the answer worse?
More thinking helps when the problem has several steps that depend on each other. It does little, and sometimes harms, when the problem is missing information rather than reasoning.
Three patterns where we would turn effort down, not up:
- The context is wrong. In our view this is the one to rule out first. If Claude has not seen the right files, extra thinking produces a longer, confident answer about the wrong code. Fix the context first; our guide to context engineering for AI agents explains how.
- The task is simple. A high level on a rename or a formatting change adds latency and cost and invites Claude to "improve" things you did not ask about.
- You are debugging by guessing. Long reasoning over a vague bug report can produce elaborate theories. A structured approach works better. Debug by Hypothesis, $5, forces falsifiable hypotheses and one read-only experiment before any fix.
We have not measured how often each pattern happens; these are the cases where, by design, more reasoning cannot add information that is not there. When answers degrade in a way you cannot explain, the testing and debugging collection is the next stop.
How do you set thinking for a whole team?
Our approach: put the default in shared settings and let people raise it per task. The Claude Code settings docs describe effortLevel as a default for models without a saved level of their own, maxEffortLevel as a cap for every model or per model and modelSettings for a saved level per model. A cap is the setting we would add first for a team, because it stops one long session at max from eating a week of budget.
A practical setup for a team of five: default medium, cap at xhigh plus a short note in the project's CLAUDE.md that says when to use ultrathink or raise the level. Pair it with a planning skill so hard tasks start in plan mode. Implementation Plan Writer, $5, turns a request into an ordered plan of small tasks with file paths and checks, which makes a lower effort level go further. Our Claude Code best practices cover the rest of the shared setup, and the prompt and skill engineering collection has skills for writing those team instructions.
Your next step
Check your current effort level with /effort, set a team default and cap and keep ultrathink for the one question a day that deserves it. If you want benchmark data in the same session when you pick models, connect your agent once through Agensi MCP. These controls changed several times in 2026, so we re-check this page monthly; check the docs for your version too.
Written by Samuel Rose, Founder of Agensi. Next scheduled review: 13-12-2026.