Every answer costs something. Most people find that out the week they run out.
What you spend, and the mechanic behind it.
Six changes. None depend on which AI you use.
Three engines. When to reach for each.
The control nobody touches.
Your seat, and running out.
Parts 1 and 2 work anywhere. Parts 3 and 4 are Claude.
Why it adds up faster than it feels like it should
Session 1 covered what a token is. Tokens are the unit your plan is metered in.
Four dials. Two of them matter.
Click your name, bottom left, then Usage.
Check it Monday, not Thursday. Note the free full reset sitting there.
You run out of the short one first.
Hit it by three o'clock. Refills on its own. Annoying, not fatal.
This is the one that ruins a week. Resets at a fixed time on your account.
Find your reset time and write it down. The top model carries its own separate weekly limit.
Every message makes it re-read the whole conversation from the top.
It has no memory between turns. The entire transcript goes back, every time.
Not a Claude thing. Every chat assistant works this way.
Which makes "thanks, that's perfect" on a sixty message thread the most expensive habit in the building.
Every file you attach stays in the conversation for the rest of it.
Open the wrong thing once, pay for it on every turn after.
Six things to change. None care which AI you use.
The whole session in four words.
People keep one chat open for weeks. That is not a relationship. It is a receipt.
Learn that second line. Say it, and the long chat is safe to close.
Five questions means five full re-reads. One message means one.
Same result. One turn instead of five.
Paste the same price sheet every morning and you pay for it every morning.
Four minutes to set up. Everyone skips it.
Read as text and as a picture of every page. You pay for both.
An image of a spreadsheet costs far more than the eight numbers in it.
The cheapest thing you can hand it, and usually the clearest.
Do you need page 31, or the whole binder? Usually page 31.
Left alone it writes long. Long costs twice: to write, then on every re-read.
The fastest saving in this deck is six words at the end of your prompt.
Every tool switched on gets described to the model before it answers.
Treat them like shop lights. On in the bay, off when you leave.
| 1 | New job, new chat | Biggest saving, by a distance |
| 2 | Right model for the job | Several times the cost, same question |
| 3 | Batch your questions | Five re-reads become one |
| 4 | Projects, not re-pasting | Stop paying daily for one file |
| 5 | Ask for shorter answers | Compounds quietly |
| 6 | Turn off unused tools | Small each time, real over a week |
One and two are most of it. The rest is tidying.
You would not run the big one to go get coffee
Quick lookups, tidying a list, pulling fields out of text.
Drafting, editing, summarizing. For most people, the only one they need.
Dense analysis, long multi step work, anything expensive to get wrong.
Start on Sonnet. Down for easy, up for hard. There is a fourth above these with its own weekly limit.
Direction, not gospel. What holds: the same chat can cost several times more on the top tier, and most chats do not need it.
| Reformat this list | Haiku |
| Pull the dates out | Haiku |
| Draft the follow up | Sonnet |
| Summarize the meeting | Sonnet |
| Why do the numbers disagree | Opus |
| Plan a change for the team | Opus |
Check what it is set to. Model and effort both carry over from last time.
Right next to the model name, and almost nobody touches it
Click the model name, then Effort.
The menu marks Medium as Recommended, and says plainly that higher effort uses your limits faster.
Routine work. You will not see the difference.
Where the menu puts you. If you are not sure, you are fine here.
Hard reasoning, long work, decisions. For one job, not as a setting.
Max warns you on the way past.
Continuing a chat is cheap because the history is cached.
Switch model or effort partway through and the cache is gone. The whole conversation gets reprocessed.
Needing a different engine halfway usually means it became a different job. New chat.
You find these settings, put the top model on max, and leave it there.
Claude even warns you. Max carries a label: 5x or more usage.
Spend the top of the range where being wrong is expensive. The bottom everywhere else.
Do not get so careful you do the work by hand to save tokens.
The goal is not using less. It is having it there on the day it matters.
| 1 | Keep them short | Every turn re-sends everything before it |
| 2 | Limit your context | Everything it reads stays in the conversation |
| 3 | Match the model | Model and effort multiply everything else |
| 4 | Preserve your cache | Switching mid chat reprocesses the lot |
Anthropic's own list. Remember the card and nothing else and you are most of the way there.
And what to do the day you run out anyway
No reason to be shy with it. Every reason not to waste it.
Hitting it every week? Say so. That is information, not a confession.
Open Usage. Put your weekly reset time in your calendar.
Your longest chat. Tell your agent to remember it, then close it.
Run a routine job on the small model at low effort. Tell the difference?
No permission, no setup. All three are settings you already have.
New job, new chat. Then match the engine to the job.
You are not trying to use less. You are trying to have it there on the day it matters.