Most people pick one AI model, throw every task at it, and call it a day. The best AI users in the world do the opposite, and it saves them a pile of money while getting better results. I came across this breakdown from an AI creator who studies how the absolute frontier of AI users work, and it flipped how I think about model choice. His core point is simple: prompt engineering matters less than deciding which tokens you use for which job.
Let me walk you through it, because it’s one of those ideas that feels obvious once you see it.
First, what a token actually is
A token is the basic unit a large language model reads and writes. Think of it like a word, roughly. The original poster showed a tokenizer where “The cow jumped over the moon” breaks into six tokens, one per word. But not always. “Programmatic” splits into two tokens, “program” and “matic.” The model works by predicting the next token, one after another.
Here’s the part that changes everything: not all tokens are equal. Some are smarter. Some are faster. Some cost way more. And the trick is matching the right kind of token to the right task.
The old way vs the smarter way
Most teams look at a new model, jump to the pricing page, and compare cost per million tokens. Sounds logical. But the expert shows why it’s a trap.
Take two models. GPT 5.6 runs about $5 per million input tokens and $30 per million output. A newer Chinese open-source model, Kimmy K3, runs $3 input and $15 output, roughly half the price. Same benchmark scores. So half price for the same intelligence, right?
Not quite. The creator pulled data showing Kimmy needs about twice as many tokens to solve the same task. So half the price times double the tokens lands you at… the same total cost. He calls this intelligence density, how much thinking gets packed into each token. In his view, OpenAI currently leads on that density. Anthropic’s Fable is powerful but burns even more tokens, which is why it topped $2.75 per task versus under a dollar for the others in the artificial analysis numbers he cited.
The lesson: cost per token is a vanity number. Cost to actually finish the task is what counts. 🎯
The three-model workflow that saves real money
Here’s where it gets practical. The mind behind this uses different models for different stages of a coding job, and the math is striking.
- Planning with the best model. He hands the hardest thinking, reading the whole codebase and writing a full spec, to a top model like Fable. This step is heavy on input tokens (reading everything) but light on output (just a plan).
- Execution with a cheap, fast model. Once the plan exists, writing the code is the easy part. He gives it to something cheaper like Grok 4.5 or Cursor’s Composer. This step is light on input but heavy on expensive output tokens, and output is where cost explodes. Remember: $30 per million output vs $5 input.
- Review with a second frontier model. He passes the finished code to a different top model, like GPT 5.6, to hunt bugs and check it against the spec. Heavy input, light output again.
The numbers he shared for one task:
- Fable alone: about $81
- GPT 5.6 alone: about $46.50
- The mixed approach: about $25.55
Same results, lowest price, and a bonus: cheap models tend to be faster, so the whole job finishes quicker too. Time is a cost most people forget to count.
Why letting two models check each other works
There’s a neat wrinkle here. Different models write and read code differently, so they catch different bugs. The creator pointed to findings from Greptile, the sponsor, showing that when Claude wrote a pull request and GPT reviewed it, GPT caught more bugs than Claude reviewing its own work. Each model is blind to its own blind spots. Teams like Nvidia, Zapier, and Substack use this cross-review approach when shipping AI-written code.
The bigger battle nobody’s watching closely
The expert closed with something worth chewing on: the price of intelligence is being decided right now. It’s closed source (OpenAI, Anthropic) versus open source (models like Kimmy).
Closed labs make fat margins selling their premium tokens, and only they can serve them, so there’s little pressure to drop prices. Open-source models are different. Anyone with a data center can host and optimize them, so every provider races to serve them cheapest. When that happens, the profit shifts away from selling tokens and flows to chip makers like Nvidia, to data centers, and to the apps built on top, because cheaper tokens mean people use way more of them.
His takeaway: if closed source wins, you keep the best tokens but pay a premium. If open source wins, prices fall and the value spreads across the whole stack.
Try it yourself
Next time you tackle a hard AI task, don’t reach for one model. Split the job: best model plans, cheap model executes, another strong model reviews. Watch your cost drop while quality holds.
This is just a primer on what the creator calls tokenomics, and there’s a lot more in the full video. Go watch it for the charts and the deeper dive.