Your Claude plan isn't too small. It's set to the wrong level.

Usually one that’s too high. People who burn through their weekly limit by Wednesday rarely ask too many questions. They ask them with a setting that is far too expensive for the job.

You don’t need a different model to fix that. With Claude Opus 5.5, the same task can cost up to eleven times as much of your plan depending on the effort level. And in most cases the more expensive result is not better.

What the effort setting does

Next to the model name, Claude has a second control: effort. It has five levels, Low, Medium, High, Xhigh and Max. It decides how long Claude thinks before answering, how often it calls tools such as web search, and how detailed the answer gets.

Opus 5.5 ships set to Medium. Most people leave it there or turn it up on principle, because more must be better. The second habit is expensive.

Effort is not a fixed budget. On a simple question, Claude thinks only briefly even on Max. A high level gets expensive on tasks with a lot to chew on, and that is exactly where it shows up in your limit.

Why your plan runs out so fast

Your plan doesn’t count messages. It counts tokens, the chunks of text Claude reads and writes. Thinking produces tokens, even though you never see them. Every level up means more thinking tokens per answer.

On top of that, chat, the desktop app and Claude Code all draw from the same pool. An afternoon of coding on a high level leaves less for chat in the evening. Anthropic’s help article on usage limits names the effort level explicitly as one of the factors that decide how fast your allowance goes.

The numbers for Opus 5.5

Anthropic doesn’t publish how much of your plan a level costs. The best approximation comes from Artificial Analysis. They run Opus 5.5 through the same tests on every level and measure what a task costs at the official API prices. Since your plan is also metered in tokens, the ratios carry over well.

I use Medium, the default, as the baseline.

Level	Usage per task	Tasks solved out of 100
Low	0.4×	42
Medium (default)	1×	51
High	1.4×	54
Xhigh	2.6×	56
Max	4.5×	58

Low and Max are a factor of 11 apart. On Max, Opus 5.5 writes around 119,000 tokens per test task on average, most of it thinking, according to Artificial Analysis. The biggest saving sits at the bottom: Low uses about 60 percent less than Medium.

What the scores mean

The right-hand column comes from the Artificial Analysis Intelligence Index, as of September 2026. It combines ten very hard tests: expert-level science questions, real office tasks with documents and spreadsheets, programming in a terminal. Roughly speaking, the number shows how many out of 100 such tasks Claude solves.

So 51 means about every second expert-level task works out. Your email, your filled-in form or your meeting notes are far easier. Every level handles those.

One caveat belongs here. Because the test tasks are so hard, Claude thinks a lot on every level. On everyday tasks the levels sit closer together in usage. The order stays the same.

Does it get better if I turn it up?

Barely. Performance rises clearly only at the start. After that you pay more and more for less and less.

Step	Extra usage	Extra tasks solved out of 100
Low → Medium	+144 %	+9
Medium → High	+36 %	+3
High → Xhigh	+90 %	+2
Xhigh → Max	+73 %	+2

High to Xhigh nearly doubles usage for two tasks out of a hundred. Those are two expert-level tasks, not two emails.

Anthropic’s own documentation on the effort parameter says Max adds significant cost for relatively small gains on most workloads, and can lead to overthinking on simple or structured tasks. The code review company CodeRabbit reached the same conclusion in its own tests: Max gave mixed results, and the default level was the better choice overall.

Important is not the same as difficult

This is the actual mistake. We pick the level by how important a task feels. The proposal for the biggest client, the contract, the tax form. That one has to be right, so High, just in case.

But a text doesn’t get more correct on High if Medium already gets it right. More thinking only helps when a task fails for lack of thinking. How much it matters to you plays no part.

The better question is: does the result have gaps? If not, the level was right. If it does, go up one level and try again.

Which level when

Level	When
Low	Enough more often than you’d think: anything whose result you can quickly check yourself
Medium	Everything else, including important and demanding work
High	Not a starting point. Only once Medium has demonstrably missed something
Xhigh	Long jobs with many steps: large research tasks, agent runs of half an hour or more
Max	Only when Xhigh has failed

Switching takes seconds: in chat through the menu next to the model name, in Claude Code with /effort low or /effort high. Getting used to a quick look at which row fits before you start saves more than any model switch. It pays off twice for recurring work, such as routine tasks AI handles in the background.

Making your allowance last

• One level up, not straight to the top. If the result doesn’t fit, try High first. The jump to Max is almost always the most expensive one with the smallest gain.
• Sharpen the request before touching the setting. A shallow answer often comes from a vague request. A line like “This has several layers, think it through carefully” often gets more out of Medium than High does with a fuzzy description.
• New chat for a new topic. With every message, Claude rereads the conversation so far. A chat with eighty messages makes every further question more expensive.
• Tools only when needed. Web search and connected apps pull a lot of text into the chat. Switch them off when you don’t need them.
• Check Max results against High. Run the same task once on High and once on Max. If you can’t find the difference, you have your answer.

The principle holds outside AI too: tuning the tool feels like progress but is often just a more expensive way of not doing the work.

Common questions

Where do I set the effort level in Claude?

In chat and the desktop app through the menu next to the model name, in Claude Code with the /effort command followed by the level. In Claude Code, Max only lasts until the end of the session.

Is Max the best choice if usage doesn’t matter?

No. Max is slower, and according to Anthropic it can overthink simple or structured tasks, which can make the result worse. Max only pays off when Xhigh has failed on a hard task.

Why doesn’t Anthropic publish exact numbers for plan usage?

Usage depends on conversation length, attached files, tools used, the model and the effort level, so Anthropic only gives relative figures between plans. The numbers in this post are an approximation based on measured cost per task.

Should I use a smaller model than Opus to save usage?

There’s no need. According to Artificial Analysis, Opus 5.5 on Low costs about as much per task as Sonnet 5 on Low and solves more tasks than Sonnet 5 on Max. Staying on Opus and choosing the level deliberately saves more than switching models.

---

Set Claude to Low for your next email and compare the result with what you get on your usual level. If you can’t spot a difference, you’ve just freed up allowance for a lot more work.