The 45% token reduction is real. What nobody measures is how much implicit context you lose along the way. I built my own benchmark and the numbers are more complicated than the headline suggests.
Long vs short prompts, tool calls vs plain text, accumulated vs summarized context. I measured the real token cost of every architectural decision in my production agents. The numbers weren't what I expected.
CodeBurn dropped on HN and forced me to calculate my real Claude Code numbers for the first time. The result was uncomfortable: on some tasks I'm burning more tokens debugging the agent than solving the actual problem. That's not a cost issue — it's a design signal.
32 years in the dev trenches. Here I write what I learned, what I broke, and what nobody tells you in the tutorials.
No spam. Unsubscribe anytime.