The Token Diet
Before you pay more, stop burning what you already pay for.
OpenAI just made ChatGPT's top plan $500 a month. From October 30, the $200 plan gets half the usage it had, 10x the Plus plan instead of 20x. Source: Business Insider. Most of us are not short on plan. We are wasting it. These are the four changes that stopped my agents wasting 1.4 billion tokens on the same plan.
1.4 billion tokens my agents never had to read.
rtk gain showed that RTK trimmed 70% of the noisy command output before my agents read it. These are test logs, file listings and git logs, not total plan usage.
tokens my agents never had to read
of noisy command output trimmed
Do this first (2 minutes, no terminal)
In Claude Code, type /model and pick Sonnet as your everyday model. Switch to Opus only when a job actually fails. In ChatGPT, use the regular model for everyday work and save the Pro and thinking models for hard problems.
Words on this page, in plain English
- Token: the unit AI reads and writes; your plan's limit is counted in tokens.
- Claude Code: Anthropic's AI assistant that works in your project folder. Get it at claude.com/claude-code.
- Terminal: the app on your Mac where you type commands. Open it with Cmd+Space, type Terminal, press Enter.
- Codex: OpenAI's coding assistant, the ChatGPT version of Claude Code.
- Hook: a small rule that runs automatically before the AI does something.
Four changes. Start with the first one today.
Filter the noise
Trim test output, git logs and file listings before the model ever reads them.
Right-size the model
The cheapest model that can do each job. Step up only when a task actually fails.
Audit your CLAUDE.md
Have Claude compare yours to my Ultimate CLAUDE.md and keep only what fits.
Route before you run
How I built MyOS Dispatch, with Jev making the routing calls, so most requests never touch the expensive model.
Filter the noise before AI reads it
Every time your agent runs a command, the whole output goes back into the model. Test runs, git logs, file listings. Most of it is noise: hundreds of passing tests, the same headers over and over, folders nobody asked about. You pay for every line of it, and it crowds out the stuff that matters.
RTK sits between your agent and the terminal and trims that output down to what the model actually needs before it reads it. You do not change how you work. It just happens.
To be clear, RTK is not my tool. It is free and open source, made by the team at rtk-ai.app. I just use it every day. It trimmed 70% of noisy command output before my agents read it.
Don't have Homebrew? Install it from brew.sh first (one command).
Installs RTK on your Mac.
Makes sure Claude Code's settings folder exists.
Installs RTK's hook into Claude Code with no questions asked.
The hook loads when Claude Code starts. Then work normally for a day or two.
Run this later to see your own savings.
brew install rtk
mkdir -p ~/.claude
rtk init -g --auto-patch
# quit and reopen Claude Code, then work normally for a day or two
rtk gainNeeds Homebrew on a Mac. Other platforms and agents (Cursor, Gemini CLI and more) are covered on the RTK site.
The cheapest model that can do each job
Running the top model on everything is like hiring a surgeon to put on a bandage. It works. It is also the fastest way to hit your limit by Wednesday.
Quick lookups, renames, formatting, summaries
e.g. Claude Haiku
Everyday coding and writing
e.g. Claude Sonnet
Architecture, hard debugging, high-stakes writing
e.g. Claude Opus
The tip: set your default to the mid model. Step up to the top model only when a task fails, then step back down. Most days you will barely touch it, and you will not miss it.
Audit your CLAUDE.md against mine
Your CLAUDE.md gets read at the start of every session, so every extra line costs you every time. But short is not the goal, right is. I keep an Ultimate CLAUDE.md that I give away. Paste this into Claude Code in your project folder and it will audit yours against mine and upgrade only what fits your business:
- ◇Rules that do not fit your business
- ◇Stale, repeated, or overly long instructions
- ◇Useful rules from the Ultimate CLAUDE.md may be missing
- ◆Only rules relevant to your business
- ◆Short startup instructions with detail files linked
- ◆Review the changes before saving
Read my CLAUDE.md (and AGENTS.md if I have one). Then read Joe Che's Ultimate CLAUDE.md at https://workshop.mastermindshq.business/giveaways/claude-md
Audit mine against his:
1. What useful rules or sections am I missing?
2. What in mine is stale, repeated, or too long?
3. What in his does not apply to my business? Skip those.
Then upgrade my CLAUDE.md. Add only what is relevant to how I actually work. Keep it short: move long details into separate docs files and point to them. Show me the changes before you save.Get the full Ultimate CLAUDE.md
Review Claude's proposed changes before saving your updated CLAUDE.md.
Put a receptionist in front of your AI
For developers: how I route every request (optional)
If you use Claude Code, you can install it from the GitHub page; if not, skip this one.
Most people send every request straight to the most expensive model. That's like having your CEO answer the phone.
So I built a receptionist. I call it MyOS Dispatch. Before any AI runs, it looks at the request and sends it to the cheapest place that can handle it:
Checking a file or a status. Handled instantly, no AI at all.
It runs the saved recipe. No AI needed.
Goes to the cheapest model that can do it, with only the files it needs.
To make those calls fast, Dispatch uses Jev, a tiny model from TypeSafe AI (not mine) that only picks from a short list of answers. It costs me about 18 cents a day.
The point: use a classifier for choices, and save the big model for real work. Picking a lane is not a job for the smartest model you pay for.
“I had an idea for an app. I got a quote for $200,000 two years ago. I built the bones of it myself in one night. Now I have it built already, and I think I've spent about $14.”
Want to learn how to do this?
Join the Business Automation Mastermind
A small, focused group of business owners who meet weekly to build real things, fast, leaving more time to serve clients and be with the people you love.
Real reactions from Business Automation Mastermind members
What members said after their sessions.
Sophia
Session 4
“We have like 55 signups for our masterclass and we've only launched maybe four days ago.”
Jenny
Session 3
“It took like 5 minutes! I just took the one which took me one year, and I said make an essence of that, and what would be the most interesting free webinar?”
Alla
Session 3
“It looks amazing, and everyone loves my website, and they can't believe that I've created it. I can't believe it myself.”
SunDari
Session 1
“I used to pay all these people; now I can do it myself.”
Ronnie
Session 2
“I already had a draft website and it created a link to my calendar, a link to my link tree, a hyperlink for my WhatsApp, all the same photos. This is like 90% of the way there to the kind of website that I want.”
Alla
Session 4
“I've been comparing myself to a bird who now has wings. I feel so free.”
Sophia
Session 4
“I was able to go in and create my full email marketing funnel. It took off so much of the work of creating the workflow.”
Johanna
Session 1
“I have particles floating in the background, a circle following my cursor, things glowing. I could not believe I built this in one session.”
Ronnie
Session 1
“It pulled in my calendar link, my WhatsApp, all my deck photos. This is 90% of the website I wanted. In 45 minutes.”
Aaqib
Session 2
“I got a functional contact form on the website using Resend. I managed to get that set up in like 20 minutes. So I'm pretty stoked.”
Pina Maria
Session 4
“I said change it and it changed my whole website in just one second. Everything worked, it was so easy.”
Johanna
Session 4
“I was so on fire and so motivated because so many pieces of the puzzle I've been working on since a long time are coming together.”
Jasmine
Session 2
“It was great. It's been something that I really want to do, so I'm super grateful. It's so easy and we can just keep building over time. It's epic.”
Aaqib
Session 3
“I did a whole target audience and persona building exercise. I can safely say that I feel addicted to Claude.”
Marina
Session 3
“It's really empowering to learn and to see, with all these different tools, what becomes possible.”
SunDari
Session 4
“Excited that I can just ask Claude anything and be guided and supported through it all.”
Quincee
Session 4
“The website's looking great. I did a brand photoshoot this past week and I feel like a brand new person digitally.”
Alla
Session 2
“I got my website online. Thank you so much, it was amazing.”
Quincee
Session 4
“I feel magnetic. I feel guided.”
P.S. Prices will keep moving. Usage limits will keep moving. The habit that holds up either way is the same one: do not pay a model to read things it does not need.