Nine Minutes for Nothing: Kids and AI Effort Settings

Last updated: August 2026

Modern AI coding tools let you choose how hard the model works, and children have no intuition at all for when that is worth doing. In July one of my students set the effort level to maximum for a task that did not need it. The generation took nine minutes, consumed a large share of her usage allowance for the window, and produced nothing better than a lower setting would have. Days later she ran a genuinely difficult task expecting it to be expensive and it used about a quarter of what she predicted.

Wrong in both directions, days apart, on the same tool. That is not carelessness. It is what happens when a control exists and nobody has ever explained what it costs.

Key Takeaways

• One maximum-effort run took nine minutes and burned heavily without improving the result.
• A hard task the same student expected to be expensive used roughly a quarter of the predicted allowance.
• Students mis-estimate in both directions, because nothing gives them feedback unless they are taught to check.
• Usage and context are two separate budgets and both need watching during a session.
• The most valuable habit is not choosing correctly, it is looking before and after.

Table of Contents

• What These Settings Actually Are
• The Nine Minute Run
• The Task That Cost Less Than Expected
• Two Budgets, Not One
• The Handoff Habit
• What I Now Teach in Order
• Frequently Asked Questions
• The Bottom Line

What These Settings Actually Are

Two separate dials matter here, and kids conflate them constantly.

The first is which model you are using. There is a ladder from small and fast to large and capable, and the larger ones consume your allowance considerably quicker. The second is the effort or thinking level, which controls how much the model rechecks and reconsiders its own work before answering. Higher effort means better results on genuinely hard problems and pure waste on easy ones. The official Claude Code documentation covers the mechanics.

The important thing for a parent to understand is that these are real spending decisions inside a fixed monthly allowance, and a child who does not know that will discover it by running out mid-project. Anthropic publishes guidance on what drives cost in these tools, and the short version is that model choice and how much the model reconsiders are the two biggest levers.

In 20 years of teaching I have never had to teach a child to budget compute before. It is genuinely new, and it is now part of the job.

The Nine Minute Run

Alice, 15, was working on an animation for her portfolio site. She had recently learned how to change the model and the effort level, which is genuinely empowering, and she did what most people do with a new dial. She turned it all the way up.

The run took nine minutes. For a 15 year old in a one hour lesson, nine minutes of watching a terminal is a long time. It also consumed a visible chunk of her allowance for that window. And the output was not better, because the task was not hard. It was a fiddly visual tweak that needed iteration, not deep reasoning.

The lesson she took from it was more useful than the animation would have been. Effort is not a quality slider you leave at maximum. It is a trade against time and allowance, and on an easy task you pay both and get nothing.

The Task That Cost Less Than Expected

The reverse case is just as instructive.

Later in the month I set her a harder task: use a strong model at high effort, sweep the entire site, and improve whatever could be improved. She expected this to be expensive, and reasonably so. It was the broadest instruction she had ever given it.

It used roughly a quarter of what she had budgeted. She had over-corrected from the nine minute experience and was now treating the strong model as something to be feared rather than something to be aimed.

This is the actual finding, and it is why I now teach this explicitly. Neither instinct was calibrated. Children do not arrive with a sense of what computation costs, and there is no feedback loop that builds one unless somebody points at the number.

Two Budgets, Not One

There are two meters running during any AI coding session and students track neither by default.

Usage is your allowance across a rolling window. When it runs out, work stops until it resets. Umar, 16, now checks his mid-session as a matter of habit. At one point in July he read 8 percent usage and 11 percent context, which told him he had plenty of room and could keep pushing.

Context is the conversation's working memory. It fills as you go, and unlike usage it degrades quality well before it runs out. My working rule with students is that results start slipping past roughly half full, and that you want to be starting fresh rather than approaching the point where the tool compresses the conversation on its own. When it does that, a large part of the detail is gone, and the student usually cannot tell which part.

The distinction matters because the fixes are opposite. If usage is low, you push harder. If context is high, you stop and start again cleanly. A student who does not separate them will do the wrong one.

Meter	What it limits	Warning sign	The right response
Usage	How much work you can do this window	Percentage climbing fast on a big model	Drop to a smaller model, or wait for the reset
Context	How well it remembers this conversation	Getting vague, forgetting earlier decisions	Write a handoff and start a fresh session

The Handoff Habit

This is the single most valuable thing I taught either student in July, and it follows directly from the context problem.

Before the conversation gets too full, you have the tool write everything important into a file: what you were doing, what got decided, what is left. Then you close the session, open a fresh one, and point it at that file. It picks up with full understanding and an empty working memory.

The handoff itself is just a plain text file in Markdown, which is worth knowing because it means there is nothing proprietary to learn. Umar picked up the whole loop in one session, creating the handoff, generating a continuation instruction, opening a new session, and confirming the new one knew everything. He connected it immediately to the note-taking system he already uses. That is a professional habit that many working developers never acquire, learned by a 16 year old in about twenty minutes.

The reason it beats letting the tool compress the conversation itself is control. You choose what survives. Automatic compression chooses for you, and it discards more than students expect.

What I Now Teach in Order

Four steps, and I do not move on until each one sticks.

1. Look at the numbers before you start. Check usage and context at the top of the session. It takes five seconds and it decides what kind of session is possible.
2. Default to the middle of the ladder. The efficient model at normal effort handles the large majority of real work. Escalate on failure, not on ambition.
3. Escalate deliberately, and time it. If a task genuinely needs the strong model at high effort, run it when your window is close to resetting so the cost lands where it hurts least. Prepare the instruction first so you are not writing it while the meter runs.
4. Hand off before you get full, not after. Write the file, start clean. This is the habit that separates a student who can run long projects from one whose sessions get progressively worse.

Frequently Asked Questions

What do AI effort or thinking settings actually do? They control how much the model rechecks and reconsiders before answering. Higher settings help on genuinely hard problems and waste both time and allowance on easy ones. In my sessions one maximum-effort run took nine minutes and produced nothing better than a normal setting would have.

Will my child run out of their AI allowance? They can, and it is a useful thing to happen once with supervision. Teaching a child to check the meter mid-session is easier than teaching them to ration blindly. Both of my students now check as a habit.

Should my child use the most powerful AI model available? No, not by default. The mid-tier model handles most real work. The strong models are for when the normal one has failed two or three times on the same problem, and they consume the allowance far quicker.

Why do AI coding sessions get worse the longer they run? Because the conversation's working memory fills up and quality degrades before it is exhausted. The fix is to write a summary file and start a fresh session rather than pushing on. My students learn this as the handoff habit.

Is this too technical for a child to manage? My students learning this are 15 and 16, and one of them picked up the full handoff cycle in a single session. Younger children can manage the checking habit with supervision. The concepts are simpler than they sound, they are just never explained.

Related Articles

• When Should Kids Start Using AI Coding Tools?, whether your child is ready for this level of control.
• Teaching Kids to Prompt AI: Why the Verb Matters Most, the other half of driving these tools well.
• Can Kids Learn AI? What Parents Need to Know in 2026, the wider picture on what AI literacy involves.

The Bottom Line

Effort and model settings are real spending decisions, and children get them wrong in both directions because nobody shows them the meter. One of my students burned nine minutes and a chunk of her allowance on a task that did not need it, then under-spent on one that did. The fix is not better judgement, it is the habit of looking. Check the numbers, default to the middle, escalate on purpose, and hand off before the conversation fills.

Want your child to actually drive these tools rather than guess at them? Book a free Discovery Call and I will show you what a session looks like when a student is managing their own budget.