Tokenomics

Over deze cursus

Auditors price work in hours. AI prices work in tokens. If you don't understand the unit, you can't judge the cost, the value, or the risk of what you're buying. This guide fixes that by letting you play with every number yourself.

Het cursusmateriaal is Engelstalig.

Wat je krijgt

• 1 PE-uren voor je PE-portfolio
• Online, in je eigen tempo
• Een eindtoets, en daarna je PE-certificaat

Modules

1. A token is a fragment of text, billed as a number

   A model never sees your words. It sees integers, each pointing to a token in a fixed vocabulary. Type below and watch your sentence turn into the thing you actually pay for.

   5 min

2. Reading is cheap, writing is expensive

   Output costs 2 to 6 times more than input on every model. That gap comes straight from how the machine works.

   5 min

3. A long chat is the same conversation, billed again and again

   A model has no memory. To stay coherent, the whole history gets re-sent and re-billed as input on every turn. Turn 20 pays to re-read turns 1 through 19. Press send and watch it happen. Then paste a "document" into the chat and watch what that does to every turn after it.

   5 min

4. Same task, wildly different bill

   Verified July 2026, and meant as examples: the market moves faster than any table. All rates per 1 million tokens, input / output. The tiers are real: premium for the most demanding judgment and deep research, frontier for judgment, mid for drafting, budget and open models for volume.

   5 min

5. Training a model and running one are different economies

   Everything in this guide so far is about inference: running a finished model. Building one is a different universe of cost, and you pay for it anyway, invisibly, inside every token.

   5 min

6. Some prices are a promotion, some are a different machine

   Three stories sit behind the price you pay. Frontier models are expensive and still sold below cost. Chinese open models are cheap because of how they are built, and they are priced below cost too. And the subscription on your own expense claim switches the meter off entirely.

   5 min

7. Does more output mean more productivity?

   Before the answer, make a guess. In a 2025 randomized trial, experienced developers used AI on real tasks. How much faster, or slower, do you think it made them?

   5 min

8. The real cost is the human who checks the output

   Cheaper generation does not make work cheaper. It moves the work: you stop making things and start checking things. Two calculators. First: how much of the promised time saving survives review? Second: what does the workflow really cost per week?

   5 min

9. Route the work like you staff a team

   Frontier models are your expensive seniors. Budget models are your tireless interns. And the best answer is often a hybrid: intern does the volume, senior or human checks the judgment. What decides between them is how much volume there is, how much judgment a task needs, and what a mistake would cost. Nine tasks, four options each. The distinctions get finer as you go.

   5 min

10. Buy intelligence on purpose

    Nine chapters, boiled down. Four rules to route your AI work by, and the ten numbers worth carrying out of this guide.

    5 min

11. What does this mean on your engagement?

    5 min

12. Final assessment

    5 min