The Real Cost of Getting Things Done with AI

A practical argument for measuring AI by the cost of reaching a completed task — not by subscription price, tokens, or prompts alone.

AI is cheap. Until it isn't.

Most of us are told to think about AI in terms of subscriptions: €20 a month, €30 a month, maybe a few extra credits if you're feeling adventurous. Nice, clean, predictable. Except that's not really what we're paying for.

A subscription gets you access. What matters is what it costs to actually get something done.

Think of an LLM like a car. The subscription is the finance payment. Credits and compute are the fuel. Your task is the journey.

And nobody sensible compares two cars only by asking which one has the cheaper monthly payment.

You care about how much fuel it burns, how far it gets you, whether it keeps taking wrong turns, how often you need to stop and correct course, and — perhaps more interestingly — whether it occasionally convinces you to drive somewhere you never intended to go.

That last part matters more than it sounds. Because AI has two very different ways of becoming expensive:

• A bad model wastes your time by getting things wrong.
• A very good model can waste your budget by being interesting.

One creates retries, corrections, and frustration. The other produces ideas, branches, follow-ups, and entirely new lines of thought that may be genuinely useful — but were never part of the original job.

Both change the real cost of getting from prompt to completion.

If LLMs Were Cars

Imagine comparing two cars like this: Car A costs €20 a month. Car B costs €25 a month. Therefore, Car A is cheaper. That tells you almost nothing. What if Car A burns twice as much fuel? What if it regularly takes the wrong exit? What if you have to stop every twenty kilometres, open the bonnet, and explain to it what a roundabout is?

The same thing happens with LLMs. A model might have a lower subscription price or look cheap on a benchmark table, but if completing a task requires five prompts instead of two, repeated corrections, re-uploading context, and checking hallucinated information, its sticker price is irrelevant.

The meaningful question is not: How expensive is the model? It is: How expensive is the journey?

The Bad Driver and the Enthusiastic Tour Guide

There are two distinct ways an LLM increases the cost of a task:

1. The Bad Driver: The answer is weak, misses context, or hallucinates. You correct it; it fixes one thing and breaks another. You spend extra prompts, checking, and manual rewriting. This cost feels like friction and waste.

2. The Enthusiastic Tour Guide: You ask it to improve a landing page. It does. Then it notices positioning could be sharper. You explore positioning, which exposes a pricing issue, which leads to redesigning the onboarding flow. An hour later, twelve tabs are open. Nothing failed — usefulness simply swallowed your boundary.

The Completion Boundary

The missing operational concept is the completion boundary: the exact point at which the task you set out to perform is sufficiently complete. Not perfect; simply good enough to send, ship, or decide.

Everything before that line belongs to the cost of the task. Everything after belongs to scope expansion. One is friction; the other is discovery. Mixing them makes real benchmarking impossible.

Cheap Intelligence, Expensive Attention

Once AI becomes cheap and permanently available, compute ceases to be the scarce resource. The scarce resource is human attention.

The cheapest AI may not be the one with the lowest price tag. Sometimes the most expensive model is the one that forces five retries; sometimes it is the one that executes cleanly and immediately offers seven compelling reasons to keep talking.