The useful result
A small score table for your own three tasks on low and high thinking, and one rule for when more thinking is worth its cost.
- Pick three tasks you can check with a yes or no list.
- Run each task on a low and a high thinking setting and score both without knowing which is which.
- Write a one-line rule for when you pay for more thinking, with your own numbers.
You hit your usage limit by lunchtime. Someone says to turn thinking up, someone else says to turn it down. Nobody has tested it on your work.
This lesson helps you test it in about 25 minutes. You will pick three tasks, run each on a low and a high thinking setting, score both by a checklist and compare the cost or limit use. If 25 minutes is too long, run one task today.
Why test? Vals AI, a company that tests AI models, studied this on 3 November 2026. On the lowest setting, Claude Opus 5.5 kept 80% of its overall score at 16% of the cost. But for Legal Research, the median gain from the lowest to the highest setting was 19.2 points. So the answer depends on the task.
Vals ran about half the tasks, mostly once, and priced them at list price, not your plan's limit. Model names and settings change, so this test works with any tool.
Before you start
- An AI tool with a choice of thinking or reasoning level. Do not paste private data.
- About 25 minutes, and three tasks that you could check by hand.
Review evidence
Documentation-based review · 2026-11-09
Sources opened and checked on 9 November 2026. Figures are dated and describe what the sources reported, not rules. Prompts and examples are illustrative; no model output or outcome is claimed.
Pick three checkable tasks and write a checklist for each
Choose three real tasks: one easy, one medium and one harder. Each result must be checkable by you with yes or no questions, such as "Is every figure from my text?" Avoid tasks where you can only say "it feels good". Use only text you are allowed to share. Write four to six checklist items for each task before you run anything. Do not change the checklist later.
Help me write a scoring checklist for each of three tasks. Do NOT do the tasks. For each task, write 4 to 6 yes/no checks that I can verify by hand from my own text, with no opinion questions. Include one check for "no invented fact".
TASK 1: [TASK ONE]
TASK 2: [TASK TWO]
TASK 3: [TASK THREE]Example / expected result
Illustrative example; not a recorded model result.
Task 1: Shorten a 120-word email to 60 words. Keep both dates.
- Is it 60 words or fewer?
- Are both dates still there and unchanged?
- Is any new fact added? Task 2: Add up a 12-row list of costs.
- Is the total right when I add it by hand? Task 3: Summarise two pages in 5 bullets.
- Does each bullet match the text? Is any figure invented?
Check your result
Each task has 4 to 6 yes/no checks, and you could mark each one without asking the AI.
Run each task on a low and a high setting
Find the thinking or reasoning choice in your tool. Setting names differ and change, so write down the exact names and today's date. Run each task once on the lowest option and once on the highest, in a new chat each time, with the same prompt. Note how long it took and what your tool shows about use, such as messages left or a usage bar. If your tool shows no choice, use this lesson on any other switch it has, but say which one in your notes. Label the answers A and B in random order so your scoring is blind.
Task: [TASK]. Use only the text below. Give the answer only, with no extra advice.
TEXT:
[PASTE YOUR TEXT]Example / expected result
Illustrative example; not a recorded model result.
Tool: any chat app, checked 9 Nov 2026. Settings used: lowest and highest on offer. Task 2, run A: total given. Took about 5 seconds. Usage bar moved a little. Task 2, run B: total given, with a check line. Took about 25 seconds. Usage bar moved more. (Made-up numbers, to show what to write down.)
Check your result
For each task you have two saved answers with hidden labels, the setting names, today's date and a note on time or usage.
Score both by your checklist and compare cost
Score A and B against the checklist for each task. Then reveal which was low and which was high. Count the yes answers. Compare the extra time or limit use. A simple rule, which is our own and not a standard: pay for high thinking only where it wins by two or more checks on a task that matters. Three tasks and one run each is a small sample, so repeat any close call once. Save the table and the date.
Below are my scores for three tasks on a low and a high thinking setting. Make a table with columns: Task, Low score, High score, Extra time or limit use, Winner. Then give one sentence per task. Do not add any number that is not in my notes. Do not say which setting is better in general.
MY SCORES:
[PASTE YOUR SCORES AND NOTES]Example / expected result
Illustrative example; not a recorded model result.
| Task | Low | High | Extra use | Winner | | Shorten email | 3 of 3 | 3 of 3 | more | tie, use low | | Add costs | 1 of 1 | 1 of 1 | more | tie, use low | | Summarise 2 pages | 3 of 5 | 5 of 5 | more | high, by 2 | On this small test, high thinking only helped on the summary.
Check your result
Your table matches your own scores. No number in it is new, and you wrote a rule that depends on the task.
When it goes wrong
Do not read these as proof that low is always enough. At the lowest setting no model kept 100% of its overall score, and higher settings helped most on Legal Research and Code Migration. If the AI changes your checklist, say: "Use my checklist exactly." If both settings look the same, repeat once before deciding. Setting names and prices change often, so retest after an update. Do not paste private or confidential data.
Make it your own
Do this today in 15 minutes: pick one task you repeat every week, write four yes/no checks, then run it once on low and once on high with the labels hidden. Count the checks each one passed.
Useful next resources
Sources and review
Last checked Nov 9, 2026. Features and account limits can change. Verify important results against your original material.
Did this help you complete your task?
Tell us what needs improving. No account or private output required.
Report incorrect information or a broken linkGenius Forges Editorial
Practical exercises based on checked sources. Examples are illustrations, not recorded model results.
