Extended and Adaptive Thinking

0:15 On every current Claude model, thinking is adaptive: Claude decides for itself whether and how long to reason before answering, and the effort level (Low, Medium, High, Extra high or Max) sets how much work it puts in (Effort, Anthropic docs, read Nov 6, 2026). In the Claude apps, click the model name next to the send button, then Effort, and pick a level (Using extended thinking, Claude Help Center, read Nov 6, 2026). The older "extended thinking", where a developer set a fixed token budget for reasoning, began with Claude 3.7 Sonnet in Feb 2025 and is no longer accepted by models from Claude Opus 4.7 on.

Quick facts

Sources: Effort, Thinking and Models overview, Anthropic docs, read Nov 6, 2026; Introducing Claude Sonnet 5.5 and Introducing Claude Mythos and Fable 5.1, Anthropic.

What is thinking in Claude?

Thinking is text Claude writes for itself before its answer: working through a problem, checking a step, planning tool calls. Anthropic introduced it as "extended thinking" with Claude 3.7 Sonnet on Feb 24, 2025, calling it the first "hybrid reasoning model" because one model could answer at once or think first (Claude 3.7 Sonnet and Claude Code, Anthropic, Feb 24, 2025). Thinking with tool use followed with Claude 4 in May 2025, and adaptive thinking, where the model chooses how much to think, with Claude Opus 4.6 on Feb 5, 2026 (Introducing Claude Opus 4.6, Anthropic, Feb 5, 2026).

What you see is not the raw reasoning. The docs say what is shown "is never the raw chain of thought": a separate model writes a summary, and on current models the API leaves thinking out of the response by default unless asked to show it (Thinking, read Nov 6, 2026). On Fable 5.1, Opus 5.5, Opus 5, Sonnet 5.5 and Fable 5, asking Claude to write out its hidden reasoning in the answer may be refused.

How do you turn on thinking in the Claude apps?

The Claude Help Center gives these steps (Using extended thinking, read Nov 6, 2026):

1. Click the model name next to the send button.
2. Click Effort and choose Low, Medium, High, Extra high or Max. The model's recommended level is marked "Default".
3. On models that allow it, hover over Effort and switch the Thinking (or Extended) toggle on or off. On older models without effort, the Extended toggle sits directly in the model menu.

Thinking is always on for Opus 5.5 and Fable 5.1, and the Help Center says it cannot be turned off in the apps on Haiku 5.5, Sonnet 5.5 or Opus 5 either. The effort selector covers Haiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.1, Opus 5, Sonnet 5, Fable 5, Opus 4.7, Opus 4.6 and Sonnet 4.6; Extra high is offered from Opus 4.7 on. Enterprise admins can hide models or effort levels. Effort came to claude.ai and Cowork on every plan with Claude Opus 4.8 on May 28, 2026 (Introducing Claude Opus 4.8, Anthropic, May 28, 2026).

Higher effort uses more of the plan's limits. Anthropic's own advice for stretching usage limits on always-thinking models such as Opus 5.5 and Sonnet 5.5 is to choose a lower effort level (How do usage and length limits work?, Claude Help Center, read Nov 6, 2026).

Effort levels

Source: Effort, Anthropic docs, read Nov 6, 2026. Opus 4.6, Sonnet 4.6 and Mythos Preview have max but not xhigh. Haiku 4.5 and Sonnet 4.5 have no effort setting.

The docs suggest starting Opus 5 and Fable 5.1 at high, Haiku 5.5 at medium, and Opus 4.7 and 4.8 at xhigh for coding, and running your own tests rather than copying settings across models: Sonnet 5.5's levels were recalibrated, so high there is not the same amount of thinking as high on Sonnet 5. More effort is not always better. Anthropic reports Sonnet 5.5 scoring lower on FrontierCode at Max than at Extra high (Introducing Claude Sonnet 5.5, Sep 28, 2026), and Artificial Analysis measured Fable 5.1 at 49 on its index at max against 50 at xhigh (Models leaderboard, read Nov 6, 2026).

Thinking on the Claude API

Set the level with output_config.effort; no beta header is needed (Effort):

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2026-10-21" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5-5",
    "max_tokens": 4096,
    "messages": [{"role": "user", "content": "Compare two database designs"}],
    "output_config": {"effort": "medium"}
  }'

The thinking field chooses the mode. Leaving it out, or sending {"type": "adaptive"}, gives adaptive thinking on current models. What each model accepts (Thinking, read Nov 6, 2026):

• Fable 5.1, Opus 5.5, Fable 5, Mythos models: adaptive only; disabled returns an error.
• Sonnet 5.5: adaptive, or between_tools, which skips thinking before the first answer, at high effort or below.
• Haiku 5.5 and Opus 5: adaptive, or disabled at high effort or below.
• Opus 4.7 and 4.8: no thinking unless you send adaptive.
• Opus 4.6 and Sonnet 4.6: adaptive, or the deprecated {"type": "enabled", "budget_tokens": N}.
• Opus 4.5, Sonnet 4.5, Haiku 4.5: only the budget form; adaptive returns an error.

Set max_tokens high at xhigh or max (the docs suggest 64,000 as a start), because it caps thinking and answer together.

What does thinking cost?

3:16 Thinking tokens are billed as output tokens, at the model's output price, which has been so since Claude 3.7 Sonnet (Claude 3.7 Sonnet and Claude Code). Thinking the API omits from the response is still billed in full (Thinking). On Claude Opus 5.5 that is $20 per million tokens, so the effort level is in practice a price setting; Artificial Analysis measured Opus 5.5 at $2.45 per task at medium and $3.46 at high on its tests (read Nov 6, 2026). See Claude API Pricing.

Thinking in Claude Code

1:27 In Claude Code, /effort sets the level (or auto), and --effort sets it at launch. Opus 5.5, Sonnet 5.5 and Haiku 5.5 default to medium there, Fable 5.1 to high (Model configuration, Claude Code docs, read Nov 6, 2026). Writing ultrathink in a prompt runs that turn at xhigh on current models (Claude Code changelog, Oct 26, 2026). The Alt+T (Option+T) toggle for extended thinking has no effect on Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable models.

What reviewers found

11:20 In a Sep 29, 2026 video, Theo Browne of Theo - t3.gg argued that effort levels work as budgets: Low to Extra high allow the model more reasoning but let it stop early on easy work, while Max raises the floor and keeps it reasoning. In his own benchmarks, going from Low to Extra high added about 5% more tokens, and going from Extra high to Max a 1,500% increase, which he said can hurt answers through overthinking. His advice for Anthropic's models was to avoid both Low and Max.

Frequently asked questions

How do I turn on extended thinking in Claude? In the Claude apps, click the model name next to the send button, hover over Effort and switch Thinking on, or on older models switch on Extended. On Claude Opus 5.5 and Fable 5.1 thinking is always on, and Effort is the only control.

What is adaptive thinking in Claude? Adaptive thinking lets Claude decide how much to reason for each request instead of using a fixed token budget. It arrived with Claude Opus 4.6 in Feb 2026 and is the only mode on Claude's newest models.

Which Claude effort level should I use? Start with the model's default, Medium on Claude Opus 5.5 and High on Fable 5.1 through the API, and raise it only when answers fall short. Higher effort costs more tokens and takes longer, and Max is not always the best scorer.

Does Claude thinking cost extra? Claude's thinking tokens are billed as output tokens on the API and count toward usage limits in the apps. There is no separate fee, but higher effort spends more.

Can I see Claude's thinking? Partly. Claude shows a summary written by another model, never the raw reasoning, and on newer models the API omits thinking unless you request it.

Is extended thinking available on the Free plan? Claude's Free plan includes Sonnet 5.5 and Haiku 5.5, both of which think adaptively, and Anthropic's effort control reached every plan in May 2026. Opus and Fable models need a paid plan.

Sources

• Effort, Anthropic docs, read Nov 6, 2026
• Thinking, Anthropic docs, read Nov 6, 2026
• Models overview, Anthropic docs, read Nov 6, 2026
• Using extended thinking, Claude Help Center, read Nov 6, 2026
• How do usage and length limits work?, Claude Help Center, read Nov 6, 2026
• Claude 3.7 Sonnet and Claude Code, Anthropic, Feb 24, 2025
• Introducing Claude Opus 4.6, Anthropic, Feb 5, 2026
• Introducing Claude Opus 4.8, Anthropic, May 28, 2026
• Introducing Claude Mythos and Fable 5.1, Anthropic, Sep 15, 2026
• Introducing Claude Sonnet 5.5, Anthropic, Sep 28, 2026
• Model configuration, Claude Code docs, read Nov 6, 2026
• Claude Code changelog, Anthropic, read Nov 6, 2026
• Models leaderboard, Artificial Analysis, read Nov 6, 2026
• Which Claude model should you use?, Claude, Sep 18, 2026 (video)
• Using Claude Opus 5.5 as your daily driver, Claude, Sep 23, 2026 (video)
• OpenAI should be scared of this one, Theo - t3.gg, Sep 29, 2026 (video)