What Does It Cost to Run?

The Horror-Story Invoice

Somewhere in every owner's feed is a story about a company that turned on AI and watched the monthly bill hit five figures. The stories are real. Almost every one traces back to the same design mistake: the most expensive model got assigned to work that never needed it, and then it ran at volume on a schedule.

Model pricing spans several orders of magnitude. Frontier models can cost dozens of times more per unit of work than small fast ones. That gap is the whole story of AI cost control. Get the matching right and runtime becomes a footnote. Get it wrong and the invoice becomes the thing your CFO wants to talk about.

Match Model Cost to Job Value

Think of models the way you think about staffing. You do not send a senior partner to stuff envelopes, and you do not send an intern to negotiate the acquisition. Models divide the same way.

Judgment work, the drafting, analysis, and decisions where a wrong answer costs real money, earns the frontier model. Assembly work, the pulling, formatting, summarizing, and cross-checking that makes up most scheduled automation, runs perfectly well on models that cost a fraction as much. Most of what businesses automate is assembly work. That is the punchline hiding in the horror stories: the expensive failures usually come from paying senior-partner rates for envelope stuffing, around the clock.

The Sorting Question

The sorting question for any automated task is simple. If this output is wrong, what does it cost me? When the answer is a minor annoyance caught by the next human in the chain, a small model is the right hire. When the answer is a bad decision, a lost client, or a compliance problem, spend up.

A few defaults that hold across most businesses:

• Scheduled, repetitive, high-volume work defaults to small fast models.
• One-off judgment work, analysis, and customer-facing drafting defaults to frontier models.
• Anything a human reviews before it matters can run cheaper than anything that acts on its own.
• Revisit the assignment when a task changes stakes, not just when the bill changes.

What Our Own Fleet Costs

Our scheduled fleet runs daily: a morning brief, an email digest, task-system syncs, a market-analysis run, research ingestion, and a supervisor agent watching all of it. Nearly every one of those runs on small fast models, because they are assembly work. The frontier models come out for judgment: analysis sessions, drafting, architecture decisions, the work where quality is the point.

We will not publish our invoice, but the shape of it matters more than the number. The fleet's monthly runtime lands at about what a single mid-tier software subscription costs, and it is a rounding error against the hours it returns every week. The interactive judgment work costs more, and it earns it, because that is where wrong answers get expensive.

How to Read an AI Line Item

When a proposal or an invoice lands, three questions sort it fast. First, which models run where? A vendor or builder who cannot tell you which tasks run on which tier has not thought hard about your money. Second, what runs on a schedule? Scheduled work multiplies daily, so a mismatched model there compounds while a mismatch in occasional work stays small. Third, what is the cost per outcome? Runtime per report generated, per lead handled, per hour of admin time returned.

A healthy AI line item reads like a payroll you would approve: cheap labor on the routine, expensive judgment where judgment pays. If it reads like everyone got the senior rate, ask why.