Why You Should Price the Task, Not the Model, Before You Automate
A new model launch always brings a headline price per token. The number that matters is the cost of your actual task.
When the Automation Bill Doesn't Match the Demo
A new AI model gets announced, the press covers its per-token price, and someone on the team forwards the link with a note: "this looks cheaper, should we switch?" Three weeks later the automation is live, the invoice arrives, and the total doesn't look anything like what the headline number suggested. This happens constantly in small and mid-sized companies rolling out their first AI-driven process — a document summarizer, a customer email triage, a data extraction pipeline.
The gap comes from a simple mix-up: the price per million tokens is not the price per task. A model can be cheaper per token and more expensive per job, because it needs more tokens to do the same work — longer prompts, more back-and-forth, verbose output that has to be trimmed or re-processed downstream. Input and output tokens are often priced differently too, and a task that reads a lot but writes little behaves nothing like one that reads little and writes pages of text.
Without a cost model built around your specific workflow, model selection turns into a guessing game driven by whichever price got the most attention that week. That's an expensive way to run a budget line that should be one of the most predictable costs in the business, because unlike headcount or rent, token consumption can actually be measured in advance.
How to Price a Task Before You Deploy
Before picking a model for an automation, run the numbers on the job itself, not on the vendor's rate card.
- Write out the actual prompt and a realistic sample output for the task, then count the input and output tokens separately — most providers publish a tokenizer or counting tool for exactly this.
- Multiply input tokens by the input rate and output tokens by the output rate; add them together to get a cost per single run, not a cost per million tokens.
- Multiply that per-run cost by your expected monthly volume — number of documents, emails, or records the process will actually touch — to get a monthly figure you can compare to a line item in the budget.
- Test the same task on two or three candidate models with real inputs from your business, not generic examples, since verbosity and reasoning length vary by model and directly change the token count.
- Add a margin for retries and edge cases — malformed inputs, timeouts, or outputs that need a second pass — because that overhead rarely shows up in a first estimate but shows up every month on the invoice.
This exercise takes an afternoon. It replaces a marketing number with a figure tied to your own documents, your own volume, and your own process — the only number that will actually appear on next month's bill.
What to Watch Once It's Running
Pricing a task before deployment is only half the job. Once the automation is live, the real test is whether the estimate holds up against actual usage.
Track the average token count per run for at least the first few weeks, not just the total spend. A rising average often means prompts are drifting — growing context, added instructions, or accumulated conversation history — long before the invoice makes it obvious. Compare cost per task against the manual cost it replaced: hours saved multiplied by a realistic hourly cost, set against the token bill plus whatever oversight the process still requires. If the automated version costs more per unit of output than a human did, the model choice or the prompt design needs revisiting, not the whole project.
It's also worth re-running the comparison periodically. Prices per token move often, sometimes sharply, as providers adjust rates or release new versions. A model that made sense six months ago may not be the cheapest option today, and the only way to know is to repeat the same measurement rather than assume the initial choice still holds.
Get Your Automation's Real Cost Checked
ArkonLabs builds automations and AI-assisted workflows for companies that want a number before they commit, not after. If you're weighing which model to use for a process, or want the real cost of a task checked against what it's replacing, reach out through www.arkon-labs.com.