Effort Levels and Extended Thinking

Learning Objectives

By the end of this lesson, you can:

• Set the effort level for a session, and read which level is in force
• Explain the difference between effort and extended thinking
• Choose a level from the shape of the task, not from a habit
• Demonstrate that thinking output is billed, using /usage
• Get one-off deeper reasoning on a single turn without changing the session

The Problem

A one-line rename took forty seconds and cost more than the previous twenty turns combined. Nothing about the task was hard.

Earlier that morning, the effort level had been turned up to max for a security review, and never turned back down. Every turn since — the renames, the typo fixes, the questions with obvious answers — had been bought at maximum reasoning price.

The opposite failure is quieter and worse. A hard bug gets a confident, wrong fix, and you conclude the model is not good enough. It was never asked to think: the level was low, or thinking was off, and a single shallow pass is exactly what you paid for.

Both failures are one dial, set by habit rather than by task.

The Concept

Two related ideas, often used as if they were one.

Effort level controls adaptive reasoning: how much the model is willing to think before it answers, decided per step from the difficulty of that step. Lower effort is faster and cheaper on easy work. Higher effort spends more on reasoning when a task rewards it.

Extended thinking is the reasoning itself — the tokens the model emits before its answer. Those tokens are real output. You are billed for all of them, whether they are shown, collapsed, or redacted.

flowchart TD
    P[Your prompt] --> E[Effort level in force]
    E --> R{How hard is this step?}
    R -->|routine| S[Little or no thinking]
    R -->|hard| T[Think first]
    T --> A[Answer]
    S --> A
    T --> B[Thinking tokens: billed as output]

The reframe: effort buys thinking and thinking is billed, so the question is never "is higher better" but "does this task need the model to think before it answers."

The levels

As of October 2026 the levels are low, medium, high, xhigh, and max, with medium or high as the common default depending on the model, and levels differ by model. A level a model does not support falls back to the highest supported level below it. Read the live table at Model configuration; the names move.

Roughly:

Two honest cautions. Higher effort shows diminishing returns, and it can overthink: more reasoning is not automatically a better answer. And the effort scale is calibrated per model, so the same level name does not mean the same amount of thinking across two models.

Turning it, and showing it

The session header shows the level next to the model name, for example "with low effort", and /effort status reports it. ultrathink is the small, useful exception: it asks for deeper reasoning on that one turn and leaves the session's level alone.

Thinking output is collapsed by default. Press Ctrl+O to expand it and watch the reasoning. On some models you cannot turn thinking off at all — the model decides per step, and effort is the only control you have.

Build It

1. Read the level in force

/effort status

Then look at the session header: the level sits next to the model name. Write down both.

✅ You pass when you can state the current level without opening a settings file, and say which control set it.

2. Make effort visible in the bill

Pick a task with a right answer — a bug fix, a tricky git command, a small script. Run it once at low, then again, identical, at high:

/effort low
... run the task ...
/usage
/effort high
... run the same task ...
/usage

Record the cost of each. If the model's thinking changed, the output-token count moved; that is where the money went.

✅ You pass when you have two costs for the same task at two levels, and can point at the token count that changed.

3. See the thinking

Run any task, then press Ctrl+O to expand the reasoning. The gray text before the answer is the thinking you are paying for.

✅ You pass when you can point at the expanded reasoning and say which bucket in /usage it landed in.

4. Ask for depth on one turn

Leave the session's level where it is, and use the keyword instead:

ultrathink: find the race condition in this file before suggesting a fix.

The effort level sent to the model is unchanged; the instruction is added for this turn only.

✅ You pass when the deeper turn happens without the session's level moving, and you can say why that matters for the rest of the session.

5. Meet the cost of changing your mind

Change the level in the middle of a conversation and watch the next turn. On most models a change of effort is a change the cache cannot help with, so the whole conversation is reprocessed. That is the next lesson's subject, and seeing it here makes it land.

✅ You pass when you have observed the slow turn after an effort change and can name the lesson that explains it.

Use It

Values and level names move — check Model configuration before quoting them.

Exercises

1. Take a task from last week that you found hard. Run it at low and at high, identical prompt, and compare both the answer and the cost. ✅ You pass when you can say whether the extra effort changed the answer at all, with two costs from /usage to back it up.

1. Find a task you do repeatedly and would never want deep reasoning for. Set the session to the lowest level that still does it well, and record the cost difference against high. ✅ You pass when the cheap level produces an answer you would ship, and you have the two costs written down.

1. Deliberately set max, run an easy task, and measure the waste. Then decide the rule you will follow for choosing a level. ✅ You pass when you can state your rule in one sentence and point to the measurement that justified it.

Key Terms

What this lesson skipped

ultracode and dynamic workflows. A toggle that has Claude plan a workflow for each substantial task. It depends on concepts from Phase 12; come back when you have them.

Per-skill and per-subagent effort. A skill or subagent can carry its own effort level. Those live in Phases 05 and 06, and they need this lesson first.

Organisation effort caps. An administrator can cap the level a model may use. Documented on the model-configuration page; not something you set as a user.

Fast mode. A separate speed setting on some plans, with its own cost trade-off. Adjacent to this lesson and covered in the model-configuration page.

How effort interacts with caching. Deliberately deferred to Lesson 22, where the mechanism explains why changing effort mid-session has a price.