AI Literacy 101: the model writes the tool, it isn't the tool
Somebody spent $321 on a 5-hour Claude session and posted the receipt calling it a scam. It's going around right now. Look at the wrong number and you'll agree with him. Look at the right one and you'll never make his mistake.
He selected Fable. Some of the work got done by Opus. He called that a silent bait-and-switch.
It wasn't. Fable hands certain queries to Opus on purpose. It's a safeguard and it's written in the docs. He wasn't scammed, he was safeguarded.
Now here's the number he skipped right past: 252 million cache reads on Opus. In one session.
What a cache read is
Every time you send another message, the model has to re-read your whole conversation to answer. Your files, your history, everything you've stacked up so far. That re-read is a cache read and you pay for it every turn.
Short session, small context, it's cheap. Sit in one session for five hours and never clear anything out, and you're paying to re-read a phone book on every single reply.
252 million reads means he never got out of the chat and never cleaned up his context. That's not the model reaching into his wallet. That's him re-reading his own mess a few thousand times and footing the bill for it.
The real mistake
A lot of people think the model is the machine. It's not the machine. It writes the machine.
You have the model write the script one time. Then you run the script. The script is code. It costs nothing to run and it does the same job every time you fire it.
The second you keep the model in the loop doing the actual work over and over, you're paying top dollar to reinvent something you already have sitting in a file. The answer's done. You're just too deep in the chat to go run it.
You probably didn't need the API at all
This whole $321 bill came off the API. Most of what people are running on the API doesn't need to be on the API.
The agentic operator knows you can do almost all of this on your subscription. Twenty bucks a month, flat. Build your workflow, write your scripts, run your terminal sessions, all under the plan you already pay for.
The API is a metered meter. Every token is a charge. That's the right tool when you're offering an AI service for someone else to use, where their usage is the product and you're billing on top of it. That's what it's for.
Running your own personal build work through the API and watching a per-token bill climb? That's a choice, and usually the wrong one. Use the subscription for your own building. Save the API for when you're the one providing the service.
Consumer, operator, builder
Consumer treats the model like the machine. Asks it to do the same task fifty times, pays fifty times.
Operator knows when a token is worth spending and when it isn't. Uses the model to think, then steps out of the way.
Builder has the model write the thing, saves it, runs it for free from then on.
The $321 guy is a consumer blaming the tool.
What to do instead
Build the workflow once with AI. Save it as a script. Run the script.
Third time you're back in the chat asking for the same kind of thing, stop. That's your cue to have it write you something you can run yourself.
And watch your session length. A long session isn't a flex, it's a cost. A clean workflow running in your terminal beats a 5-hour chat on price and on control, every time.
The point of using AI is so you can stop using it for the parts a script already handles. If your bill looks like his, you never made that jump.
Just Keep Building.
1
2 comments
AI Literacy 101: the model writes the tool, it isn't the tool
skool.com/cmdandconquer
CMD & Conquer — stop buying AI tools and learn to build the things being sold to you. Become the operator who directs AI. Terminal-first.
Powered by