Writing2 min read
Checkpoint: two weeks in, four billion tokens later
Why I haven't posted, and what it's like to go from rationing free tokens to burning billions of them.
I haven't posted here in a while. Part of the reason is the internship I started on October 23rd. Only part, though: I wasn't posting much before it either. But since the 23rd I have barely found time for anything at all. It's called an internship, but at almost nine hours a day it's a job in everything but name.
You think you can handle it. Everyone does. Then it hits you like a truck, and there was no way to see it coming.
So this isn't a real post. It's a checkpoint, a note to future me about where I was.
From rationing tokens to burning billions
Until this month, almost all the AI I used was free or close to it. I did use Claude Opus 5 quite a lot, but through cheap third-party routers like AgentRouter, not Anthropic directly, and my lifetime usage, priced at API rates, would come to maybe $500. Most of the time I was on DeepSeek and the other free models OpenCode provides, and I got stuck with them more often than I'd like to admit.
A long time ago I got a month of OpenCode Go for free. It was one of the best months I've had as a developer, because I got to try a lot of models. I was also so careful with the quota that by the end of the month there was a big chunk left and no way to use it up in time. I wanted to finish it and couldn't. That's how conservative I was with tokens.
The internship comes with Claude Code. The numbers so far:
- About 1.5 billion tokens in roughly the first six days
- Almost 2 billion tokens on November 4th alone
- Another 600 million today
That's over four billion tokens in two weeks, almost all of it on Claude Opus 5.5, from someone who once couldn't finish a free monthly quota on purpose. It's thrilling. Opus 5.5 is a genuinely very good model.
The internship
Honestly, Claude Code is the part of this internship I care about most.
The company's roots aren't in tech, and the culture still shows it. It isn't the engineer-first kind of place you read about online. Then there are the access politics. You'd think a startup would hand you access to basically everything. Instead, even the smallest thing means finding the one person who controls it and asking, politely, whether they could please give you access to this one too. I could talk about it all day, but not here.
The rest is fine. Nothing glamorous, nothing terrible. I'm enjoying it about as much as I'm enjoying it.
Now I'm going to leave something running overnight in Claude Code. More when I find the time.
Update, November 8
That 600 million turned out to be a fraction of the day. My usage for the four days since November 4th:
| day | tokens |
|---|---|
| November 4 | 2.02 billion |
| November 5 | 5.06 billion |
| November 6 | 2.20 billion |
| November 7 | 5.70 billion |
That's almost 15 billion tokens in four days. To be fair to my past self, about 96% of it is cache reads, the same context read back on every turn. Actual output was around 55 million tokens. Still. lol.