1.5 billion! - That’s the approximate amount of tokens I spent on Claude Code, in January! And the number is conservative, for it excludes usage across 4 other machines in the cloud. Not included here is a fair bit of 2x Cursor subscriptions and Codex subscriptions. I recently added Gemini to the repertoire for evals. GitHub Copilot was my first ever sub. They’ve clearly lost the plot, so I plan to vote with my wallet.
Claude /insights tells me I spent 2500 hours sending approximately 11,000 messages, over the course of a month (1500h and 1800 messages for work projects and 1000h and 9661 messages on personal), churning out an aggregate of +77,226/-30,291 (no-slop) LOC and +998,800/-357,983 (partial-slop) LOC. An additional 10k-20k or so were spat out by cursor and codex, as a footnote.
The numbers are just incomprehensible to me. That’s over 50x my generous estimate of 20k LOC. That is more code than I’ve written in many years. 2500 hours! At 10-15+ hours a day, my brain split across 5 parallel agents at any given moment. Often the number was higher, because I did eat and sleep! No wonder I am beat.
I’ve also reviewed a surprising amount of code, where on occasions, I reckon I spent more time reading than the author did writing. Am now able to spot the specific model and tool an author chose to use, just by reading the code.
Iain M. Banks’s Culture series left me believing AI and superintelligence will let the human mind work less. Yet I find myself working twice as much. We are fast approaching a point where the meatspace I live in is the bottleneck.
Where do we go from here?