ClaudeExtended thinkingLevel 3: Workflows

When Should You Turn Up Claude's Thinking?

On current Claude models, you don't decide whether Claude thinks. Claude decides on each request, and the control you have is effort. Raise effort for multi-step problems where a wrong answer costs you, and lower it for quick work you can easily check. The old Extended thinking mode, with a fixed token budget, is now a legacy feature.

Prompt to copy
Setting: click the model name next to the send button, choose Effort, and pick High or Extra high for this chat.

Then send:

This needs careful reasoning.

Problem: [describe the decision or problem]
Constraints: [list each one]
What I already tried: [attempts and what happened]

Give me your recommendation first, then the two or three facts it depends on. Before you finish, verify your answer against each constraint and tell me which one is tightest.
  1. lowquick work you can check
  2. mediumdefault on Opus 5.5, Haiku 5.5
  3. highdefault on most other models
  4. xhighlong agentic and coding work
  5. maxhighest capability
Effort levels, lowest to highest. Source: Anthropic's effort docs, checked 6 November 2026.
↑ cost of a wrong answer
Default effortsimple, but check the result
High or abovehard and costly to get wrong
Lowquick rewrites and lookups
Default effortmany steps, easy to redo
steps of reasoning →
A rule of thumb, not an Anthropic chart: leave effort at the default, and move it only when the task sits at a corner.

Try it now

  1. 1Open a chat and pick a real problem you've been stuck on.
  2. 2Click the model name next to the send button, choose Effort, and note which level is marked Default.
  3. 3Send the problem at the default level, then open a new chat, choose one level higher, and send the same prompt.
  4. 4Compare the two answers against what you know. If you can't name a real difference, keep the lower level. The Thinking section above each response shows a summary of the approach.

Use it when

Raise effort for an architecture decision, a bug that has survived two attempts, a contract clause that interacts with three others, or a plan with competing constraints. It also suits any answer you'd act on without being able to check it easily.

Skip it when

Skip it for short rewrites, formatting, translation and lookups, where you can see at a glance whether the result is right. Higher effort there only costs you time and usage. If you're close to your weekly limit, try the default first and raise effort only if the answer falls short.

Why it works

Anthropic's docs call this adaptive thinking: Claude judges each request, so a simple factual question may get no thinking at all while a multi-step math problem or tricky debugging task gets deep reasoning. Effort sets how readily and how deeply it thinks. The levels are low, medium, high, xhigh and max. The default is medium on Opus 5.5 and Haiku 5.5, and high on most other models. Effort applies to every output token, not only thinking, so higher levels take longer and reach your plan's limits faster. On the API, thinking tokens are billed as output.

Example

Say you're choosing how to migrate 40,000 customer records to a new billing system without a downtime window, and you have to keep invoice numbers stable. That has several dependent steps, and a mistake reaches customers, so it earns High effort. Asking Claude to shorten your announcement email about the migration doesn't, so switch back down for that.

Going further

In Claude Code, phrases like think hard are ordinary prompt text. The Claude Code docs say effort is the real control, and ultrathink asks for deeper reasoning on one turn without changing your setting. On the API, set output_config.effort. Changing the top-level value between requests invalidates prompt cache breakpoints, so pick one for a conversation and use per-message effort (beta) where a single turn needs more.

Common mistakes

  • Hunting for an Extended thinking switch on Opus 5.5 or Fable 5.1. Anthropic's help center says thinking is always on there and the toggle doesn't appear, so lower effort instead.
  • Sending budget_tokens from an old code sample. Anthropic's docs say Claude 4.7 and later models reject it with a 400 error.
  • Leaving Max on for every chat, then running out of your plan's limits on routine work.
  • Expecting high effort to guarantee a thinking block. Anthropic's docs say no level does, and follow-up requests in a tool loop can skip it.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course