Price 1 Finished Task, Not 1 Million Tokens: The Only AI ROI Number an Owner Needs

NVIDIA counts tokens. That is the right unit for a company selling computing power. It is the wrong unit for pricing work in a small firm.

A price per million tokens tells you what the raw material costs. It does not tell you what one finished, checked, usable piece of work costs your firm. Those two numbers can point in opposite directions: a task can be cheap in tokens and expensive in the attention of the people who prompt it, check it, correct it, and decide what to do with it.

What a token is, in plain terms

In "What Are AI Tokens?" (NVIDIA Blog, Dave Salvator, 17 March 2025), NVIDIA explains that a token is a small unit of data made by breaking larger information down. A word such as "darkness" can be split into "dark" and "ness", and that shared "ness" helps a model relate it to "brightness". The same process applies beyond text to images, audio, and other data.

Three parts matter to an owner. Input tokens are what you send in; output tokens are what the model writes back. The context window is how many tokens it can process at once: a few thousand is roughly a few pages, while some models now offer a million or more. Newer reasoning models also generate extra reasoning tokens while they work. NVIDIA states that this can require over 100 times more compute than a single pass by a traditional model — a statement about compute, not a promise that your bill will rise by the same multiple.

NVIDIA is also open about its goal: the fastest processing and the lowest cost per token. It calls data centres "AI factories" manufacturing intelligence. That is vendor framing from a company that sells compute — sensible at infrastructure scale, but not the goal of a ten-person firm trying to clear inboxes without creating a second checking job.

Why the per-million price misleads owners

Our companion guide to what AI really costs a small firm sets out token prices, seat prices, and a calculator in full, so this article will not repeat those tables. It uses the same verified worked examples, at vendor prices checked on 12 October 2026, to make a different point.

On a mid-range model at $2 per million input tokens and $10 per million output tokens, summarising 100 emails costs about $0.24 in tokens. That sounds trivial, and at token level it is. Now consider an agent working through a task in 20 steps, re-reading a 30,000-token context at every step. That is 600,000 input tokens for one task, about $1.20 in tokens, and around $240 a month if it runs 200 times. Still not a large bill — until you add the part that never appears on a token invoice.

Who prompted it, checked it closely enough to trust it, corrected it, and made the final call? If an experienced person spends 15 or 20 minutes doing that on each run, checking can cost many times the token charge. A cheaper model needing three attempts and a rewrite can cost more per finished task than a pricier model that gets it right first time.

Tokens are the spend unit. Attention is the budget

The useful bargain is simple: let the machine absorb routine work — first drafts, summaries, sorting, standard replies — so scarce attention stays on judgment. That idea has older roots in people and computers working in symbiosis, and in the modern case for protecting deep, focused work. Spend tokens freely on routine work; guard the minutes of the people whose judgment clients pay for.

A workflow that saves $0.10 in tokens while consuming 20 minutes of senior checking has not saved money. It has moved cost from a small, visible invoice to a larger, invisible one.

The same trap appears with felt speed. In the METR study, experienced developers believed AI had made them faster while timed work told a different story. We covered that in developers felt faster with AI, but timed work was slower. Feelings, seat counts, and token prices are proxies. Finished work is the measure.

Price one finished task, end to end

Pick one repeated task with a clear definition of "done": not "AI drafted something", but work you would send to a client, file, or act on without embarrassment. Then price it both ways for a week.

Cost per finished task equals token and seat costs for that task, plus prompting, checking, correcting, and deciding minutes multiplied by the rate of the person doing each part, divided by the number of tasks that met your standard. Count retries. Count escalations. If half the output needs a rewrite, those tasks are not finished at draft stage.

Compare that with the same finished task before AI. Our AI ROI page shows how to frame that comparison before you buy. If AI wins on finished work at the quality you require, expand it. If it only wins on token price or felt speed, you have learned something cheap. Set a kill date in advance, as in escaping AI pilot purgatory with a 90-day proof test, so a weak tool does not drift into a permanent subscription.

Illustrative worksheet — your documents, model, and minutes will differ: runs per month × token cost per run, plus checking minutes per run × that person's hourly rate, divided by finished tasks that passed. Leave the checking line blank and the worksheet will lie to you in the vendor's favour.

Frequently asked questions

How much do AI tokens cost?

It depends on the model and whether tokens are input or output, with output usually costing more. At vendor prices checked on 12 October 2026, a mid-range model at $2 per million input and $10 per million output tokens summarised 100 emails for about $0.24. Token tables and a calculator are in our guide to what AI really costs a small firm. That charge alone is not the cost of the finished task.

How much does AI cost per month?

The monthly cost is seats plus usage plus staff time. Seats are fixed per person, paid whether used or not. Usage tools add token charges — about $240 a month for 200 runs in the example above, before checking time. Price per finished task first, then multiply by how many finished tasks you actually need each month.

Should I choose the model with the lowest cost per token?

Not on that figure alone. Choose the workflow with the lowest cost per finished task at your quality standard. A low token price can be offset by retries and corrections; a higher one can be cheaper if it finishes in one pass with only a light review.

What should I measure before renewing an AI tool?

One repeated task, timed before and after, counting only finished work that met your standard, including every prompting, checking, and deciding minute. If you cannot name that cost, renew for the shortest term while you measure it — or do not renew.

Start with one task

Start with one weekly task, define "finished", and price it honestly for seven days.

Tell Ada the one task you'd automate first. She'll price it as a finished task — tokens, seats, and checking time — in your free report. You can talk to Ada and request your free report — no raw token table will make the decision for you, but one finished-task price can.

Sources: NVIDIA Blog, "What Are AI Tokens?", Dave Salvator, 17 March 2025, for tokens, input/output, context windows, and reasoning tokens. NVIDIA sells compute; its "AI factories" framing is vendor framing, and no NVIDIA savings claim is relied on. Dollar examples are from BuildBrain's costs guide at vendor prices checked on 12 October 2026. The worksheet is illustrative.