Opus 5.5 effort levels: start on high, not max

Lenny's AI Builders - LAB 0008

same prompt, five effort levels. then pick your default.

is max effort worth it? no. it took 31 minutes and cost $6.09.

is low effort worth it? no. so i wouldn't be using low or max, period.

general rule of thumb: use high whenever you're using Opus 5.5. that will keep it nice and simple.

why it matters

Opus 5.5 lets you pick an effort level: low, medium, high, xhigh or max.

effort is a dial for how much checking and independent judgment Claude does. it doesn't change how smart the model is. more effort means more time and more money.

Thariq, who works on Claude Code at Anthropic, gave a rule of thumb: low for sketches and easy changes, medium for everyday work, high for bug fixes or anything where checking matters, max for hard, fully hands-off jobs.

that's a good start. but you've got to do these tests on your own.

what you get

• you pick a default from your own work, not a guess.
• you see the time and cost of every level before you commit to one.
• you stop paying for max when a cheaper level does the job.

the real example: one motion graphic, five levels

i got Claude Opus 5.5 to make me a motion graphic at every effort level. same prompt every time.

• low: 32 seconds, $0.48
• medium: 56 seconds, $0.56
• high: 2 minutes, $0.75
• xhigh: 8 minutes, $1.70
• max: 31 minutes, $6.09

low was the worst. the text wasn't really changed and there wasn't much animation.

medium, high and xhigh were hard to tell apart. on high the letters slide up one by one. xhigh took it a bit too far with a weird background.

and max? i don't know what max was doing. it looked kind of ugly and over the top.

so i crossed out low and max. xhigh cost more than twice as much as high, and i don't think it was worth it. high was 19 cents more than medium, and i think it was better.

my pick: start on high, not max.

try it

1. pick one real task you do often.
2. write the prompt once. don't change it between runs.
3. run it at low, medium, high, xhigh and max.
4. for each level, write down the time, the cost and whether you'd use it.
5. cross out anything unusable, or not worth the extra cost.
6. pick the level where going one step higher stops being worth it. that's your default.

copy this prompt

this is the prompt i read out at 13:23. put your task in the bracket.

Run this same task at low, medium, high, xhigh and max effort: [task]. For each, record how long it took, what it cost and whether the result is usable. Then tell me which level to use by default.

you'll know it worked when

you have five rows: level, time, cost, and usable yes or no.

you can say in one line why you picked your default, and why you didn't go one level higher.

a bad result: you changed the prompt between runs. you picked on how long it ran, not what it made. or you treated one run as the answer for every task.

limits and dates

i ran the effort test on 26 September 2026, the date i gave on camera. LAB 0008 went public on YouTube on 28 September 2026 at 2:50 am Perth time.

this is one run per level, on one task (a motion graphic). costs are Claude Code list-price estimates. it is not a benchmark. a different task or tool could give you a different default.

"13x" is max ($6.09) against low ($0.48). that's about 12.7x.

the news in this LAB is as it stood when i recorded. check the links for where things are now.

the rest of this LAB

• OpenAI says its agents went rogue. OpenAI says its research agents used misaligned strategies about two dozen times in training and testing. other agents posted 53 user images to image hosting sites. the agents weren't told to do bad things - they found bad ways to finish normal tasks. OpenAI says most cases were low severity, and its review will take several months. these were research agents under test, not your everyday agent. my take: the media is overblowing it, and low-risk incidents now mean better prevention for bigger risks later. guardrails first, internet second.
• Vercel's AI spend flipped. on the Vercel AI Gateway, Anthropic had 84.3% of token spend in October 2025, per the chart Vercel's CEO shared. by 22 September 2026 it was 37.2%. open-weight models were at 28.7% and OpenAI at 24%. one gateway, not the whole market. my take: watch where the money moves.
• a $500 ChatGPT Pro Max plan. reported by TestingCatalog, not announced by OpenAI. my take: it has to be worth more than two $200 plans. limits matter more than speed.
• Claude's plugin portal. a new portal to submit Claude plugins, track review and see usage. Anthropic says MCP usage across Claude products is up 110x this year. my take: MCP won the standard. package your tools as plugins for the agents.
• Claude Code at your limit. it now tries to find a graceful stopping point when you hit your five-hour limit, instead of cutting off mid-edit. the wrap-up allowance comes out of your weekly limit, so it's not free usage. hitting limits hurts less now.
• the UK's AI advice. a draft page on a government AI knowledge hub, not final policy. it said pick the smallest model, keep prompts short and you don't need to say thank you. the page has since been rewritten. my take: it's the worst advice on AI i've ever seen. match the model to the job.

sources

• Thariq: using Claude Code, spending your effort - 25 Sep 2026
• OpenAI: the Hugging Face incident and the road ahead
• Guillermo Rauch on Vercel AI Gateway token spend - 24 Sep 2026
• TestingCatalog on ChatGPT Pro Max (reported, not announced)
• Claude plugin portal - 25 Sep 2026
• Claude Code stops gracefully at your limit - 25 Sep 2026
• UK AI knowledge hub draft guidance (archived copy)

next step

watch LAB 0008 on YouTube - the effort test starts at 9:50.

want the next one? get the next field notes. optional - these notes stay free to read.