Nobody comes home and announces they only used 0.45 gallons

Return from the shops and nobody says "we only used 0.45 gallons, result". They ask whether you got there, whether you remembered the milk, and whether the car is still in one piece. Fuel is a cost of achieving something, not the thing itself, and treating it as the objective would be regarded as a peculiar way to live.

Somewhere in the adoption of this technology we inverted that completely. Token consumption became the number people quote, optimise and boast about, while whether the output was any good got filed under secondary considerations. It is fuel economy on a route that was wrong.

Cheap attempts add up quietly#

The trap is that each individual attempt looks inexpensive, which is exactly what makes the total invisible.

You ask a question. Five thousand tokens, negligible. The answer is close but pitched wrong, so you rephrase. Another five thousand. It has assumed context you did not supply, so you clarify. Five more. It is now technically correct and useless for your purpose, so you regenerate with different framing. By the time you have something you can use you have spent twenty five thousand tokens and forty minutes of a life that does not come with a refund.

Every one of those attempts was cheap. The sequence was not. And nobody records the sequence, because each step felt like a small correction rather than a cost, which is precisely how expensive habits establish themselves.

We normalised trial and error because the unit price was low, and then built entire practices around it: prompt libraries, internal wikis of what works, training in the dark art of asking. All of that effort exists to compress necessary context into something small enough to feel efficient. It is solving the wrong problem with impressive discipline.

What single-shot actually asks for#

Look at what you are requesting when you fire one prompt and hope.

You are asking a system to correctly infer your intent, your expertise level, the format you need, the audience, the constraints you did not mention, and the standard you will judge it by, from a paragraph. That is not prompting. That is mind reading, and then we express surprise when the mind reading is imperfect.

The alternative is not more artful prompting. It is defining the context properly and stopping. Who you are, what you are trying to achieve, what shape the output must take, who receives it, what would make it wrong. That costs more up front, and it keeps paying, because every subsequent exchange builds on a foundation instead of reconstructing it from scratch.

Infrastructure rather than rent. You pay once and stop paying the re-explanation tax on every interaction.

The multi-agent bill looks worse and is not#

The same logic scales up, and it looks even more alarming on the invoice.

Instead of one generalist handling everything, you orchestrate specialists. Something gathers material. Something analyses it. Something checks consistency against the requirement. Something assembles the result. Each step consumes context and produces output, and the total makes a token-conscious person wince.

Then count what disappeared. The three rounds of clarification. The two regenerations. The manual verification a person performed because nobody trusted the output. And, most valuably, the variance: run it again next week and you get comparable results rather than whatever the model felt like that morning.

Consistency is worth more than efficiency in almost every business process, and it is the thing single-shot prompting structurally cannot deliver. A process you can rely on is a process you can build on. A process that is excellent on Tuesday and mediocre on Thursday is not a process, it is a talented individual with an attendance problem.

Cost per activity#

The reframing that changes the conversation with anyone holding a budget has nothing to do with tokens.

Express it as cost per completed activity. Take a task somebody performs manually, work out what it costs when a person does it including the fully loaded cost of that person's time, and set it against what it costs to do automatically. In one orchestration I worked on, the answer came out at a few pence per activity for something that had taken a person about ten minutes. On its own that sounds like nothing. Multiply by several thousand occurrences a month and the picture assembles itself without anyone needing to explain what a token is.

No finance director has ever asked about token consumption. They ask what it costs to do the thing, and how reliably it gets done. Those are answerable questions, and answering them in their language ends the argument in a way that no efficiency metric ever will.

What the obsession actually protects#

There is something worth naming underneath all this. Optimising tokens is measurable, immediate and entirely within the control of whoever is doing it. Optimising outcomes requires agreeing what good looks like, which means a harder conversation with people who may not have decided.

So the token metric persists partly because it is the thing you can improve without asking anyone's permission. I have some sympathy. I have also watched teams spend months shaving consumption on a workflow that was producing results nobody used, which is a perfectly efficient way to achieve nothing.

The question is not how few tokens you can use. It is what level of investment produces output you can depend on, and whether the fully loaded cost of the humans currently correcting things exceeds the difference, which it almost always does by a wide margin.

Nobody will remember how efficiently you failed to deliver what was needed.