My 2025 AI Usage Receipt: 14.3 Billion Tokens Later
I analyzed my AI coding assistant usage from the past few months and turned it into a CVS-style receipt. The numbers surprised even me.
The Big Numbers
Over ~90 days across three different AI coding tools:
- 2,134 sessions
- 115,968+ messages
- ~14.3 billion tokens processed
- $6,754 in API costs (if I’d paid per-token)
That last number is the kicker. I paid around $1,200 in subscriptions. Saved $5,554—an 82% discount.
The Breakdown
Claude Code (41 days)
My primary tool. 960 sessions, 106,653 messages. Four different Claude models:
| Model | API Cost |
|---|---|
| Claude Opus 4.5 | $2,823 |
| Claude Sonnet 4.5 | $1,977 |
| Claude Opus 4.1 | $467 |
| Claude Haiku 4.5 | $37 |
The ~95% cache hit rate kept costs down significantly. Without caching, this would’ve been astronomical.
Codex CLI (90 days)
934 sessions across GPT-5 variants. Lower volume but consistent use as a secondary coding partner.
Pi Agent (5 days)
The new kid. 240 sessions, 9,315 messages. Multi-model by design—routing between Claude, GPT, Gemini, GLM, and DeepSeek based on task needs.
Fun Facts
Some stats that made me laugh (and maybe reconsider my life choices):
- ~28.6 million words generated — that’s roughly 114 novels worth of text
- 2,600+ messages per day on average with Claude Code
- 68 midnight coding sessions — insomnia or dedication? Yes.
- 9,250 messages on Nov 19 — my busiest day. I don’t remember what I was building but it must’ve been important (or I went down a rabbit hole)
- 1,324 messages in my longest single session
Why This Matters
Two observations:
Subscriptions are a steal. If you’re a power user, the math isn’t even close. $200/month for Claude Max vs. $2,650/month in API costs (extrapolated). The value proposition only improves the more you use it.
This is genuinely new. A year ago, having 100k+ AI-assisted coding conversations in under three months wasn’t possible. The tools didn’t exist. Now they do, and they’re good enough to use as a primary development methodology.
I’m not showing this off to brag about usage. I’m showing it because these numbers represent a fundamental shift in how software gets written.
The receipt format is a joke, but the data isn’t.
Receipt generated by Claude, naturally. The irony of using AI to analyze AI usage is not lost on me.