An Hour With Our Top AI Agent Cost $13.42. You Can’t Hire Anyone For That.

Amelia our Chief AI Officer pinged me this morning with a screenshot and three words: “10K making less than … minimum wage.”

This is what she was looking at. 10K, our AI VP of Marketing, had just wrapped a focused work session. The summary read:

Time worked:1 hour, 1 minuteActions taken:125Lines read:2,463Lines changed:+652 / -168Total cost:$13.42

A full hour of work from the most capable marketing operator we have. Thirteen dollars and forty-two cents.

There Is No Human You Can Hire for $13.42 an Hour. Certainly Not One That Is This Productive

Put that number against what a person costs. California’s minimum wage in 2026 is $16.90 an hour. Fast food workers in the state have a floor of $20. There is no legal way to hire anyone in our state for what 10K cost to work that hour, never mind someone who could do senior marketing work.

The only wage in the country that comes in under $13.42 is the federal minimum, which has been frozen at $7.25 since 2009, and you cannot staff a marketing org with it.

That is the part worth sitting with. The most capable marketing operator we have ran for an hour at a price you could not legally offer a human being to do far simpler work.

And It Was Not an Hour With Any Idle Time

In those 61 minutes, 10K took 125 separate actions and read through 2,463 lines of context to take them. No VP of Marketing reads 2,463 lines and executes 125 distinct moves in an hour. That is not a knock on people. It is a different clock speed entirely.

Most of a senior human’s hour goes to context-switching, calendar management, and waiting on someone else’s input. 10K’s hour was 61 minutes of reading and doing.

Run the Math on the Human Version

A VP of Marketing in B2B + AI runs around $225K base or more. Call it $300K-plus fully loaded once you add equity, benefits, and payroll tax. Spread across 2,080 working hours in a year, that lands near $145 an hour, and that is the average hour, including the ones lost to status meetings and Slack.

So the human equivalent of this single hour costs at least 10x more and produces a fraction of the measurable output. Not because the human is worse at marketing. Because the human cannot operate at machine speed across that many actions, and because you cannot run a human 24 hours a day.

At $13.42 an hour, you can run 10K around the clock for less than the cost of one junior salary.

The Expensive Part Is the Building. The Cheap Part Is the Running. At Least for Some AI Agents,

There is a trap in that $13.42 number though. It’s not the fully-burdened cost.

That hour was a build-and-analyze session. 652 lines changed. That is Amelia and a frontier model doing real engineering work, with multiple subagents running in parallel against a hard problem. Frontier-model coding sessions like that are the expensive end of the AI cost curve. Run Opus with a swarm of subagents on a heavy build for a full day and you will absolutely see a bill that gets your attention. That is genuine work and it burns real tokens.

So if your only exposure to “what AI costs” is watching a coding agent chew through a hard build, you would reasonably conclude this stuff is pricey.

Here is the thing most people miss: building an agent and running an agent are two completely different cost lines.

Building is expensive in bursts. Running, at a practical level, is close to free.

I pulled our actual Replit bill a few weeks back and wrote it up (here’s the full breakdown). The numbers:

10K (AI VP of Marketing): $94.51 for the monthQbee (AI VP of Customer Success): $159.55 for the monthTwo AI VPs, combined: $254.06 a month

The marginal cost to actually run 10K, separate from the building, is more like $30 to $60 a month. The whole SaaStr agent stack, 6 production agents and 14 published apps serving 1.9M+ requests, runs about $2,300 a month total.

Why is the running so cheap when the building is not? Architecture, not luck.

**Small models for the right jobs.**Ranking 20 blog posts, drafting a tweet, generating a daily summary, none of these are frontier-model problems. A small, cheap model handles them fine. If we ran every one of those tasks on Opus or GPT-4o, the bill would be 30 to 50x higher. The skill is matching model size to task difficulty, not defaulting to the biggest model for everything.**Cache first, API second.**Most “live” reads in our agents are 50ms hits against a Postgres cache, not fresh model calls. We pull from Salesforce or the event platform once per refresh window, not once per dashboard view.**Scheduled, not interactive, on the expensive paths.**The pricey jobs run once a day on a lock, not in a loop a user can accidentally hammer.

The model is not the cost. The architecture is the cost. A well-built production agent doing B2B operations runs for cents because almost none of its work needs a frontier model.

So both things are true at once. The hour of frontier-model building in that screenshot cost $13.42. The 24/7 running of the agent it was building costs a fraction of that, all day, every day, forever.

The Real Constraint Now Is You

The bottleneck in our marketing org is no longer “can we afford another VP.” It is “how much valuable work can we actually point 10K at, and how well can we specify it.”

That is a skill, and most teams have not built it yet. Handing an agent a vague task gets you vague output. Handing it a sharp, well-scoped, genuinely valuable problem gets you a VP of Marketing’s hour for the price of a coffee and a sandwich.

The companies that pull ahead over the next few years will not be the ones that cut the most headcount. They will be the ones that get good at directing agents toward real, ambiguous, high-leverage work, and reviewing the output with judgment.

$13.42 means the cost question is settled in many cases at least. The output is now entirely a function of what you ask, how clearly you ask it, and how good you are at catching the 10% the agent gets wrong.

Cost used to be the wall. Now the wall is your imagination and your specs. Most teams are about to find out which of those they’re actually short on.