Price per million tokens is a vendor number. It doesn’t help you decide anything. The question leadership needs answered is different: what does it cost to complete a whole task — process an invoice, answer a complaint, prepare a proposal — and what did it cost to do it by hand?
Tokeneconomics means translating model usage into that unit.
From price per token to cost per task
A task is rarely a single call. It’s usually a chain: read a document, extract data, query a system, draft, review. Each step consumes input and output, and some repeat when the first attempt isn’t good enough.
You calculate cost per task by adding up the whole chain, retries included, and dividing by the tasks actually completed — not the ones attempted. A system with a 30% retry rate costs considerably more than the optimistic estimate suggests.
With that figure in hand, you can make the comparison that matters: cost per automated task versus cost per manual task, including the time of the person who reviews it.
The levers that move the bill
Context caching. When many requests share a large, identical block — a manual, a catalog, long instructions — paying for it on every call is throwing money away. Context caching charges for it once and reuses it at a reduced price for a while. It’s the lever with the best effort-to-savings ratio.
Batch processing. Anything that doesn’t need an immediate answer — sorting overnight email, summarizing documentation, enriching a catalog — can be sent in batches at a notable discount in exchange for waiting. Much of a company’s workload isn’t interactive but gets billed as if it were.
Prompt size. Instructions grow on their own: every incident adds a sentence and nobody ever removes anything. Reviewing and pruning them lowers the cost of every call at once.
Long context windows. Putting an entire case file into the context is convenient and sometimes the right call, but you pay for it on every request. Compared with retrieving only what’s relevant through RAG, long context wins for one-off deep analysis and loses at high volume.
Routing. Sending each task to the model that fits it is, in most cases, the lever with the most room to run. It has its own page.
Budget as a business decision
An AI system with no spending cap is an open-ended risk. The reasonable minimum: a maximum cost per run, a maximum cost per day, an alert when average cost per task drifts from plan, and a named owner for that budget line.
Cost per task also marks the point where a use case stops making sense. If automating costs more than doing it by hand, the right answer is not to automate — and it’s better to know that before you build.
When it’s worth digging into
Yes, when:
- You already have one or more use cases in production and the bill grows every month
- You’re deciding whether to scale a pilot: cost per task is the number that backs the decision
- You need to defend AI investment to a committee with numbers, not impressions
No, when:
- You’re at proof-of-concept stage and spend is marginal: validate the value first
- Volume is so low that any optimization saves less than it costs to implement
How we approach it
We instrument first: every call tagged by use case, with its cost and outcome. After two or three weeks of real data, the breakdown emerges — usually a handful of operations account for most of the spend — and that’s where we act: caching, batching, prompt pruning, and routing. We measure the result against the previous month’s bill at the same volume and quality.
Frequently asked questions
How much can you save without hurting quality? It depends on your starting point. Where nothing has been optimized, caching and batching usually deliver the first big cut without changing the model or the output. The honest number only shows up after measuring.
Doesn’t the vendor take care of this? The vendor provides the mechanisms — caching, batching, model sizes — but deciding where to apply them is architecture work, and it depends on your tasks.
How do you present this to the leadership team? With two numbers per use case: cost per completed task and savings versus the manual process. Everything else is technical detail.