Your model bill is lying

I picked one model and threw everything at it for a year

July 27, 2026

In partnership with

That was my whole strategy. Open the same chat, paste the task, whatever it was. Research, code, editing, a birthday message. One model, every job, no thought behind it.

Then I hit a breakdown from someone who studies how the top one percent of AI users actually work, and it made my approach look lazy in a way I couldn't argue with. Their point wasn't about prompts. It was about tokens: which ones you spend, and on what.

That reframing did more for my AI bill than any prompt template I've ever saved. Let me walk you through it, because it's one of those ideas that feels obvious the second you see it.

Win AI Search Without a Big Team

92% of VCs use AI to find companies. 58% of buyers start there, too. If you're not showing up in AI answers, you're invisible before the conversation even starts. Join HubSpot for Startups, Anthropic, and Marketing Against the Grain on July 30 (11 am ET) for a live AEO teardown. Real startup. Real recs. Register and unlock the free Startup Visibility Bundle.

*Ad

First, what a token actually is

A token is the basic unit a model reads and writes. Roughly a word, but not always.

They showed a tokenizer where "The cow jumped over the moon" splits cleanly into six tokens, one per word. Then "programmatic" splits into two, program and matic. The model works by predicting the next token, over and over, until the answer exists.

Here's the part that changes the math: not all tokens are equal. Some are smarter. Some are faster. Some cost a lot more. The whole game is matching the right kind of token to the right kind of work.

The pricing page is where good decisions go to die

Most teams evaluate a new model the same way. Open the pricing page, compare cost per million tokens, pick the cheaper one. Sounds like basic diligence. It's a trap.

Take two models side by side. GPT 5.6 runs roughly $5 per million input tokens and $30 per million output. Kimmy K3, a newer Chinese open-source model, runs about $3 input and $15 output. Half the price, same benchmark scores. Free money, right?

Not quite. The data they pulled showed Kimmy needs about twice as many tokens to solve the same task. Half the price times double the tokens lands you back at the same total bill.

They call this intelligence density: how much thinking gets packed into each token. By that measure OpenAI currently leads. Anthropic's Fable is genuinely powerful but burns even more tokens getting there, which is how it topped $2.75 per task in the artificial analysis numbers they cited while the others came in under a dollar.

So the number on the pricing page is a vanity metric. The number that matters is what it costs to actually finish the task.

Own AI deployment, grow your career

Making AI actually work day to day is becoming its own job. Hear from three people doing it: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. They share what the role really looks like, how it came to exist, the skills worth hiring for, and the challenges they're tackling right now. Watch the full conversation on demand.

*Ad

The three-model workflow, and the receipts

This is where it stopped being theory for me. They run different models at different stages of a single coding job, and the split follows where tokens are cheap versus expensive.

Plan with the best model. Hand the hardest thinking to a frontier model like Fable: read the whole codebase, write a full spec. This stage is heavy on input tokens (it reads everything) and light on output (it writes one plan). Input is the cheap side.

Execute with a fast, cheap model. Once a real spec exists, writing the code is the easy part. Give it to something cheaper like Grok 4.5 or Cursor's Composer. This stage is light on input and heavy on output, and output is where the bill explodes. Remember: $30 per million out versus $5 in.

Review with a second frontier model. Pass the finished code to a different top model, say GPT 5.6, to hunt bugs and check it against the spec. Heavy input, light output again. Cheap side, high value.

The numbers they shared for one task:

  • Fable alone: about $81

  • GPT 5.6 alone: about $46.50

  • The mixed approach: about $25.55

Same result, a third of the cost. And a bonus most people forget to count: cheap models are usually faster, so the job finishes sooner too. Time is a cost line nobody puts on the spreadsheet.

Why two models checking each other beats one model checking itself

There's a wrinkle here I keep thinking about.

Different models write and read code differently, so they miss different things. They pointed to findings from Greptile showing that when Claude wrote a pull request and GPT reviewed it, GPT caught more bugs than Claude reviewing its own work.

Every model is blind to its own blind spots. Same reason you can proofread your own writing four times and still miss the typo a friend spots in eight seconds. Teams at Nvidia, Zapier, and Substack ship AI-written code through exactly this kind of cross-review.

37 Free Claude Prompts With The AI Report

Subscribe to The AI Report, the free 5-minute daily AI brief for 400,000+ business leaders, and you’ll get 37 Claude prompts free in your welcome email. They’re organised by the 8 situations every manager faces. You get both: the newsletter and the prompts.

*Ad

The fight over what intelligence costs

They closed with something bigger, and it's the part I've been chewing on since.

The price of intelligence is being set right now, and it's closed source (OpenAI, Anthropic) against open source (models like Kimmy).

Closed labs earn fat margins on premium tokens, and only they can serve those tokens. No competitive pressure, no reason to drop the price. Open-source models work differently. Anyone with a data center can host and optimize them, so every provider races to serve the same model cheapest.

If that race wins, profit stops living in the tokens. It flows to the chip makers, the data centers, and the apps built on top, because cheap tokens mean people burn far more of them.

Their read: if closed source wins, you keep the best tokens and pay a premium for them. If open source wins, prices fall and the value spreads across the whole stack.

Try it on your next hard task

Don't reach for one model. Split the job.

Best model plans. Cheap model executes. A different strong model reviews. Watch the cost drop while the quality holds.

This is only a primer on what they call tokenomics, and the full video has the charts and the deeper dive. Worth the watch.

The thing I got wrong for a year

I spent that year tuning prompts. Rewriting instructions, testing phrasings, saving templates, chasing a few percent.

Meanwhile the real lever was sitting one level up, in a decision I was making without noticing: which model gets which piece of the work. Prompt quality is a dial. Model routing is a switch.

Tonight, take the next real task on your list and cut it into three. Let the expensive model think, the cheap one type, and a third one argue with the result. Then check what it cost you.