I'd been using Claude Code daily for two months before I looked at the bill. It was higher than I expected — and I had no idea which sessions were worth it.
So I started tracking. Here's what 30 days of real Claude Code usage actually looks like.
The setup
I use Claude Code primarily for TypeScript and React work — feature development, debugging, refactoring. A typical session runs 45–90 minutes. I track everything through Operon, which captures token counts, cost, and tool usage per session automatically.
The numbers
- 47 sessions — about 1.6 per day
- $142 total — $3.02 average per session
- 2.8M tokens — ~60K per session on average
- Top 3 most expensive sessions: $18.40, $14.20, $11.80
The expensive sessions weren't the most productive ones. The $18 session was a 3-hour rabbit hole that ended in a revert. The $2 sessions were often the clearest, most focused runs.
Where the money goes
- Claude Sonnet 4.6 — 78% of sessions, 61% of cost
- Claude Opus 4.7 — 11% of sessions, 34% of cost
- Claude Haiku 4.5 — 11% of sessions, 5% of cost
Opus sessions cost roughly 6× more than Sonnet sessions but don't always deliver 6× better results for frontend work.
The most expensive mistake
The loop problem. Three times in 30 days I had sessions where Claude Code got stuck — editing the same file repeatedly, or running 8+ consecutive search operations without making progress. Those sessions averaged $12.40 each.
Once I started seeing these in the activity trace (Operon flags them as loop patterns), I started catching them early and resetting the session. The next month: no runaway sessions.
What this means for your workflow
If you're not tracking cost per session, you're flying blind. The average is manageable ($3/session), but the outliers can be painful — and more importantly, they're usually a signal that the session went sideways.
The most useful thing isn't the total number. It's the cost-per-session breakdown, which shows you which types of tasks are expensive for AI and which aren't. For me: debugging sessions are cheap and effective. Large refactors are expensive and need careful scoping upfront.