I thought I was a heavy Claude user.

I thought I was a heavy Claude user.

I'm spending 3 to 4 hours a day building an internal tool with Claude Code, and over the last 30 days, I burned through about 25 million tokens. That felt like a lot. Then, I compared notes with my principal architect.

1.5 billion tokens.

Same 30 days. 60x my usage.

I was very excited. Once the shock wore off, I realized this variance is exactly what we need to figure out.

Some of it is obvious. I'm a part-time user building a side project in between meetings and at night when everyone is asleep at home. And, of course, I have a leadership gig to worry about. My principal architect lives in the terminal all day. Our use cases and working patterns are completely different.

But it's the non-obvious gaps that are fascinating. Where does that 60x actually come from at the task level? What is he getting out of that volume that I'm missing?

I don't have the answers yet, but it made one thing incredibly clear: you can't build a realistic AI cost or adoption model for an engineering org from the outside.

The variance between individuals and roles is just too massive for top-down estimates to mean anything. You have to get into the weeds, look at the actual telemetry, and figure out what the super-users are doing differently.

That one conversation changed how I'm thinking about our AI strategy.