When max effort pays off, and when it doesn't

Max effort pays off on a bug that survived two tries, a costly design choice or many edge cases. For small, clear work a lower level does the job.

Contents

Max effort is worth it when the task is hard to get right, such as a bug that survived two attempts, a design choice that is expensive to undo or a change with many edge cases. For renames, small edits and everyday work with a clear scope, a lower level does the same job on fewer tokens. Start at the default and raise it with /effort when an answer comes back wrong.

What the level changes#

The effort level sets how much the model reasons before it acts. Anthropic's model configuration docs say each level "trades token spend against capability". They describe max as the level for hard problems, and add that it "may show diminishing returns and is prone to overthinking".

That thinking is not free. Anthropic's cost docs say thinking tokens are billed as output tokens. On a Pro or Max plan they count against your 5-hour window and weekly limit like any other output, and on an API key they are on the bill.

Codex has the same dial, model_reasoning_effort, which sets how much the model reasons where the model supports it (OpenAI's config basics). Everything below holds for it too, and what each Codex plan allows is in Codex usage limits.

When the top levels pay off#

We raise the level when a wrong answer would cost more than the extra thinking. That covers four kinds of work.

  • A bug that survived two attempts at the default. The first answers missed something, and the next turn needs a deeper read of the code.
  • Design and architecture, where a wrong early choice is paid for in every later turn.
  • Changes with many edge cases: migrations, concurrency, code that handles money or permissions.
  • A substantial task you would otherwise plan step by step yourself, which is what Ultra Code is for. It is a Claude Code setting that plans a workflow for each substantial task and sets effort to xhigh while it is on.

When a lower level does the same job#

  • Renames, small edits and "what does this do" questions. You read the result anyway.
  • Everyday features with a clear scope. Anthropic's docs name medium for exactly this.
  • Follow-ups in a chat that already has the context, like "now add a test" or "fix that lint error".

Anthropic's first suggestion for simpler tasks is a lower effort level, and on a subscription /usage can recommend a change such as "Switch from max to high effort". Our effort levels guide lists every level and how to set it.

A simple way to decide#

  1. Leave new chats on the default level.
  2. When an answer comes back wrong or shallow, raise the level for that chat (/effort high, /effort xhigh or /effort max), or put ultrathink in that one prompt.
  3. When the hard part is done, bring the level back down.

Anthropic doesn't publish how many tokens each level uses, and we haven't measured it, so your own usage is the best guide; on a subscription, /usage shows it. The model matters as much as the effort: Sonnet suits well-scoped work, and Opus long or open-ended work (Opus vs Sonnet for coding).

In Shuriken#

Each chat has a Thinking picker beside the model, so one hard chat can run on Max while the rest of the project stays on the default.

Shuriken counts the turns each device runs on Extra high, Max or Ultra Code. It shows one small notice, "You've been running a lot of turns on Max." (or the level you used), only when the last 14 days hold at least 40 such turns, spread over at least 4 days, and they make up at least half of all turns. It shows at most once a week and never in your first week. Don't show again turns it off; Settings › Notifications › Hidden tips and notices lists it as Thinking suggestion, with Show again.

Sources

  1. Claude Code docs: Model configuration (accessed 22 October 2026)
  2. Claude Code docs: Manage costs effectively (accessed 22 October 2026)
  3. OpenAI: Codex config basics (accessed 23 October 2026)

Frequently asked questions

Is max effort always better?

No. Anthropic describes max as the level for hard problems and warns that it may show diminishing returns and is prone to overthinking. For simpler tasks, Anthropic's own advice is to lower the effort level.

Does higher effort use more of my limits?

It can. Thinking tokens are billed as output tokens, so more thinking is more output counted against a Pro or Max plan's limits, or more on an API bill.

Is Ultra Code an effort level?

Not quite. It is a Claude Code setting that plans a workflow for each substantial task, and it sets effort to xhigh while it is on.

Free tools