I Spent $47,000 on Claude Code in 90 Days. My CTO Asked Me One Question and I Couldn’t Answer It.

The bill came in on a Friday.

I was in the kitchen at the office refilling coffee when our finance lead walked over and asked if I had a minute. He was holding a printout. Finance people in 2026 do not walk up to engineers holding printouts unless something is on fire.

The number on the printout was $47,213. Three months of API spend. One team. Eight engineers.

I told him it was fine. Velocity was up. Story points were up. The CTO had cited us in the last all-hands as an example of "AI-leveraged execution." I had a slide somewhere with a chart that bent up and to the right.

He nodded, said okay, and walked away. He did not look reassured.

The next Monday I was in the CTO's office.

The setup

We bought into Claude Code in January. The pitch made sense. Our roadmap was packed, hiring was frozen, and the agentic coding wave was real. I had benchmarked it on three internal projects in December. The numbers were good. Not magic. Good enough that I signed off without losing sleep.

By April we had eight engineers using it daily. Most of them had stopped writing functions from scratch. The good ones used it as leverage. The okay ones used it as a crutch. The new hires used it as a teacher, which is the worst of the three but I will get to that later.

Our sprint velocity numbers looked like a company that had quietly doubled headcount. Senior leadership loved the graphs. I loved the graphs. The graphs were the problem.

The meeting

The CTO did not open with the bill.

He opened by asking how the team was doing. I gave him the usual update. Shipping on schedule. Two big projects ahead of plan. A junior who was crushing it. He nodded through all of it.

Then he said: "If we paused our Claude Code spend tomorrow, how long until our velocity matched what it was in December?"

I opened my mouth to answer. Then I closed it.

I tried again. I started saying something about team adjustment periods. He waited. I could hear the HVAC.

The honest answer was that I did not know. I had not measured it. I had not built a single piece of internal infrastructure that worked independently of the agent. I had not asked any of my engineers to ship a non-trivial feature without it in six weeks.

I gave him a number. I said four to six weeks of recalibration. He wrote it down. He did not believe me. I did not believe me either.

The follow-up question

He let the four-to-six-weeks answer sit. Then he asked the question I have been thinking about ever since.

"Devrim, if the price doubles next quarter, what's our move?"

This is the question nobody on engineering Twitter is talking about. Everyone is arguing about whether agents replace engineers. Nobody is talking about what happens when the agent becomes the dependency and your entire roadmap has been quietly priced against a vendor's gross margin assumptions.

Claude Code at $47K a quarter was a line item. Claude Code at $94K a quarter was a hiring decision. Claude Code at $188K a quarter was a strategic crisis.

I did not have a contingency plan. I had not negotiated a cap with the vendor. I had not built any internal tooling that would let us downshift gracefully to a cheaper model. I had not stress-tested the team's ability to function on raw Cursor without the high-end inference loop.

What I had was a velocity chart. Velocity charts are what engineering leaders show their CTOs when they do not understand what they actually built.

The thing nobody admits

After the meeting I went back to my desk and pulled the audit logs.

What I found was worse than the bill.

The spend was distributed unevenly. Two engineers were responsible for 61% of it. One was a senior engineer I respected enormously. The other was a new hire who had been with us for eleven weeks.

The senior engineer was using the agent for code review and architecture exploration. Heavy reasoning loops. Long context windows. The kind of usage where you are actually delegating thinking, not typing.

The new hire was using it to write code he could not have written on his own. Not in a "stretching beyond his level" way. In a "he genuinely did not understand the code that landed in our repo with his name on it" way.

I had been measuring his velocity. I had not been measuring his comprehension.

When I asked him to walk me through a function he had shipped two weeks earlier, he stared at it for ninety seconds and said it was probably from one of the refactor PRs the agent had done. He was not lying. He was telling me the truth about a code review process I had let collapse.

This is the cost of agent-augmented engineering nobody puts on the invoice. It does not show up in the API bill. It shows up six months later, when you have a $94K quarterly spend and three engineers on payroll who cannot ship without it.

The Turn

I went home that Friday and could not sleep.

What kept me awake was something more uncomfortable than a budget line. I had spent twelve years building a craft, and I had failed to notice the moment I stopped applying that craft to my own team.

I was managing a chart. I had stopped managing engineers. I had stopped managing code quality. I had stopped asking the boring, slow, unsexy questions that senior engineering leaders are paid to ask. Questions like: can my team rebuild this without the vendor? Can my newest engineer explain the last thing he shipped? What is our blast radius if the price changes?

The agent did not make me a worse engineering manager. The agent made it possible for me to be a worse engineering manager and have my dashboards still look great.

That is the gap. That is the thing the bill was actually measuring.

What I did next

I did not cancel Claude Code. That would have been theater.

I did three things instead.

I capped per-seat usage at a number that forced engineers to be deliberate. I instituted a "ship one feature per sprint without the agent" rule for everyone, including me. And I made code walkthroughs part of our weekly review, where the engineer who shipped the code had to explain it to a peer with the editor closed.

Three months later our spend is down 38%. Our velocity is down 11%. Our incident count is down 64%.

I do not know yet if those numbers are causal. I know they are mine.

What I wish I had before that meeting

If you are running an engineering team that has quietly become agent-dependent, you probably do not need to read what I wrote next. You will find out at your own version of that Friday meeting.

If you want the operating system I should have been running on my team before the bill arrived, before the new hire question, before the velocity chart became a crutch, I built it into Production Engineering OS. It is the slow, unsexy set of habits that lets a senior engineer answer the question I could not answer.

You probably will not need it. Until your CTO walks into your office holding a printout.

I write the longer versions of these stories and the postmortems behind them at my Substack. Subscribe if the printout in your finance lead's hand feels personal.