Claude Extended Thinking: When It Helps and When It Wastes Usage

Where the Controls Are Now

The model menu sits next to the send button and shows the current model and effort level. Inside it you pick the model, set the effort, and, on models that allow it, toggle thinking. Changes apply from Claude's next response, so you can switch mid-conversation.

Anthropic's help center lists five effort levels: Low, Medium, High, Extra high, and Max, with each model's default marked in the menu. Extra high is available on Opus 4.7 and newer. Thinking and effort are separate settings: thinking decides whether Claude reasons in a visible section first; effort decides how thorough the response is.

The important 2026 change: thinking cannot be turned off in the Claude app for Haiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.1, or Opus 5. If you are on one of those, the question is no longer "thinking on or off?" but "how much effort?"

What Thinking Costs

Thinking is generated text. Even when you only see a short summary in the expandable "Thinking" section, Claude produced reasoning to get there. Anthropic states plainly that higher effort levels use more tokens, so you reach usage limits faster, and that lower effort stretches usage further on routine tasks.

The cost shows up as time as well. A high-effort response on a hard problem can take noticeably longer, with a timer showing while Claude works. For a two-line email reply, that wait buys nothing.

It also compounds with everything else that drives usage. Anthropic's usage best-practices article lists conversation length, attachment size, and effort level among the factors that affect limits. A high-effort turn in a short chat is one thing; the same effort on turn 40 of a chat with three large PDFs attached is much more expensive, because the model is reasoning over all of that context. If you find yourself raising effort late in a long conversation, that is usually the moment to summarize and continue in a fresh chat at the higher effort instead.

None of this means thinking is wasteful. On the right problem, one high-effort answer that is correct beats three low-effort answers you have to argue with, both in time and in total usage. The waste comes from applying it uniformly.

Decision Table: Which Effort for Which Task

Anthropic's own guidance is that the defaults work well for everyday tasks. Raise effort when the problem is genuinely multi-step, not because the topic feels important.

A Worked Example: One Session, Two Effort Levels

Imagine a 90-minute work block with 25 messages: 18 are small edits and quick questions, 7 are a hard data-pipeline bug. Running all 25 at Max effort spends deep reasoning on 18 messages that did not need it. Running all 25 at Low risks shallow answers on the 7 that did.

The better pattern is to switch. Keep the default for the small stuff, raise effort when you hit the bug, and drop back afterwards. Because changes apply from the next response, this costs one click each way. You also get better bug answers by giving the high-effort turn everything it needs in one message instead of drip-feeding it:

[Effort: High]
The nightly import fails about 1 in 5 runs. Facts:
- Error: "duplicate key value violates unique constraint orders_pkey"
- Only on runs where the upstream file has > 10k rows
- We retry on timeout; retries were added last week
- Relevant code is pasted below

List the 3 most likely causes ranked by probability, the evidence
for each, and the single cheapest check that would rule each out.
Don't propose a fix until we've confirmed the cause.

The last line matters. At high effort, Claude will happily reason its way to a confident fix; asking for ranked hypotheses and checks keeps the reasoning pointed at what you can verify.

When More Thinking Makes Things Worse

• Underspecified prompts. More effort on a vague request produces a longer, more confident guess. Fix the prompt before raising effort.
• Style and tone work. Voice edits rarely improve with deeper reasoning, and over-thought rewrites can drift from what you asked.
• Already-long chats. High effort on top of a huge history multiplies cost. Summarize into a new chat first.
• Questions that need facts, not reasoning. If the answer depends on a current price or policy, web search or a source document helps more than effort.

Reading the Thinking Section

When thinking is on, Claude shows a "Thinking" indicator with a timer while it works, then an expandable "Thinking" section above the response containing a summary of its reasoning. Most people never open it. It is worth opening in three situations:

• The answer surprised you. The summary often shows which assumption Claude made. If it assumed the wrong database, the wrong audience, or the wrong deadline, correct that assumption rather than asking for a rewrite.
• You are about to act on it. For a plan or a diagnosis, check whether Claude considered the alternative you were worried about.
• Answers keep missing. If the reasoning keeps circling the same ambiguity, the prompt is the problem, and more effort will not fix it.

Two caveats. The section is a summary, not a full transcript of the model's reasoning, so do not treat it as an audit log. And Anthropic notes that thinking can occasionally stop early for safety reasons, in which case a message says the rest of the thought process is not available; the response still arrives.

A quick way to calibrate effort for a type of task you do often: run the same well-specified prompt once at the default and once at High, and compare the two answers. If you cannot tell which is which, the default is enough for that task, and you have just saved yourself the higher cost on every future run. If the High answer catches something real (an edge case, a wrong assumption, a missed constraint), you have found a task that deserves the extra effort.

A Simple Effort Routine

1. Leave the model's default effort on for general work.
2. Drop to Low or Medium for batches of mechanical edits.
3. Raise to High for problems with several plausible answers.
4. Use Extra high or Max sparingly, for one well-specified hard question at a time.
5. Watch your session usage; if you are close to the limit, finish the hard question first and do the easy work at lower effort.

On Enterprise, admins can restrict certain models or effort levels by role, so not every option may appear for you.

ClaudeKit Workflow Note

Effort decisions are easier when you can see where you stand. ClaudeKit's usage circle in the bottom-right of claude.ai shows session and weekly limits with reset times on hover, whenever Claude returns that data. If you are at 85% of your session with a hard problem left, that is the moment to save Max effort for the one question that needs it. The live token counter also shows estimates on Claude's responses, which helps you notice when answers are running far longer than the task needs.

For timing hard work around resets, read how to plan work around Claude's 5-hour window.