$29 per merged pull request. $53,000 in tokens. Here’s exactly how we counted it.
Nobody publishes their AI costs. Most companies treat spend like a black box. Money goes in, “productivity” comes out, and nobody asks what the ratio actually is.
We’re going to be specific, including the denominator.
The $29 number is simple: $53,000 divided by 1,832 merged pull requests. That’s the unit cost from the same telemetry window as the spend.
It is not $53,000 divided by 4,106.
The 4,106 number is our Q2 team total: merged pull requests from quarter start to quarter end, across four engineers, with the CEO excluded. We verified it two ways against GitHub. During that quarter, the team averaged 79.3 merged PRs per developer per week against a floor of 75. Median cycle time was 2.4 hours, with 88% completed in under 24 hours.
Tells us what a merge costs in tokens.
Tells us what the factory produced in a closed quarter.
Those two numbers are measuring different things. $29 tells us what a merge costs in tokens. 4,106 tells us what the factory produced in a closed quarter. Mixing the denominators is how this claim dies in the comments.
And if we’re going to publish the number, we should also be clear about what’s included in it.
The $53,000 is token spend. It does not include seats, laptops, salaries, or the year we spent building the machines. We’re not presenting $29 as a fully loaded cost per pull request because it isn’t one.
$53,000, from the same measurement window as the 1,832 merged PRs.
Seats, laptops, salaries, or the year we spent building the machines.
What we are publishing is the number nobody else seems willing to put next to their volume: how much we spent on tokens and what that spend produced over the same measurement window.
That distinction matters because AI spend without output doesn’t tell you much. Saying a team spent $10,000 or $100,000 on AI tools sounds meaningful until you ask what happened on the other side of that spend. Did people actually use the tools? Did more work ship? Did cycle time change? Did the increased capacity translate into anything useful?
Cost only becomes interesting when you can put it next to output.
Linear has talked publicly about its use of Claude Code. With 45 engineers, their reported output works out to about 13 merged PRs per engineer per week. We’re at 75.
Four engineers. Q2 2026.
45 engineers, from their reported output.
Named peer. Same play.
We’ll make that comparison because there’s something concrete to compare. What we won’t do is quote an “industry average” just to make our own number look impressive. Every benchmark we’ve tried to verify has fallen apart once we looked closely enough at how it was calculated.
If the denominator isn’t clear, the comparison isn’t useful.
The question isn’t whether $29 is cheap. On its own, that number doesn’t tell you enough to decide whether it’s cheap or expensive.
The more interesting question is whether you can name your number at all.
How much are you spending on AI? What output are you measuring against that spend? Are the two numbers from the same period? Can you trace them back to the source? And if somebody challenges the denominator, can you explain exactly how you calculated it?
We can.
That’s why we’re publishing the $29. It’s not a trophy. If AI is becoming part of how engineering teams operate, its cost shouldn’t live in a black box.
We’re hiring a security leader, an engineering leader, and a VP of Sales. Have your AI send what you built in the last 30 days.
Have your AI send us what you built in the last 30 days.
Chad leads CampusIQ’s move to agentic engineering. Four engineers now merge about 1,150 production PRs a month, roughly the output of a 50-person team, with agents writing most of the code and engineers owning the outcomes, the quality bar, and the calls. He built the machine behind it: a standing merge queue, a fleet of estimation and review bots, executor-validator quality gates, and live telemetry on velocity and AI cost. Last month, more than 20 people across sales, ops, product, and onboarding shipped production code on the same system.
Aaron Benz : Sep 9, 2026, 11:32:34 AM
We set out to see if four engineers, working with AI, could produce like a team five times the size. What we built along the way changed who can...
Chad Walters : Oct 8, 2026, 10:00:00 AM
I was an engineering manager. I ran teams. I hadn't written production code in fifteen years. Now I ship more merged pull requests per week than any...
Chad Walters : Sep 2, 2026, 4:24:15 PM
I wrote this in November. I had no proof. Three weeks later I learned I was right: the best people for AI-first development aren't engineers —...
Chad Walters