I Paid $30 to Build an HTML File
Earlier today I asked Claude to help me build a simple HTML dashboard. Team planning for the quarter: who works on what, when. Pure HTML, no frameworks, maybe 200 lines of code. Should've cost a dollar or two.
It cost $30.40 in API-equivalent terms. I'm on the Max plan, so I didn't actually pay that. What I paid was two hours without Claude. The first time in seven months.
Honestly? That hurt more.
What Actually Happened
The task was simple: I had a team, a handful of projects, and a few two-week blocks across the quarter. I wanted a nice visual grid. So I opened Claude and started talking.
The problem is that I talked the way humans talk. One thing at a time.
"Put Alice on Project X." -> Claude edits the file. "Actually, make it full-time." -> Claude edits the file. "Add Bob too." -> Claude edits the file. "Next project: Charlie at 50%..." -> Claude edits the file.
Seventy-nine times. Claude edited a 15KB HTML file 79 times, re-uploading the entire thing on each pass. The file was accessed 96 times across a 226-turn session that lasted 3 hours and 14 minutes.
17.8 million tokens. For a file I could've built in Notion in 20 minutes.
Oh, and Anthropic is currently running a promotion with 2x usage during off-peak hours. I did this during peak hours. Just to round things out.
Why It Got So Expensive
Two things compounded on each other:
1. Micro-edits instead of batching. I gave Claude one data point at a time and it acted on each one immediately. Every edit re-uploaded the full file to the API. 79 edits x 15KB = ~1.7MB of the same HTML file sent over and over.
2. Cache thrashing. Claude uses a prompt cache to avoid re-processing the same context. Each edit reset the cache boundary, forcing it to be rebuilt. The session burned 4.9 million cache creation tokens (the expensive kind) instead of reading from cache cheaply.
The actual conversation? 409 real input tokens. The other 17.4 million were overhead from the way I worked.
What I Should Have Done
Collect all your data first, then generate once. Instead of "add this person, now add that one," I should've given Claude the full picture upfront (all the people, all the projects, all the allocations) and asked for the HTML in a single shot. That's maybe 1-2 edits total, not 79.
Batch your changes. If you do need to iterate, send multiple changes in one message. "Update these 5 things" is one edit. "Update this thing" five times is five edits, five cache resets, five file re-uploads.
Think in outputs, not conversations. For generation tasks (HTML, code, documents), Claude works best when you treat it like a compiler: give it complete input, get complete output. Not a live editor you nudge one line at a time.
The dashboard looks great, by the way.
And after two hours staring at the "you've hit your limit" screen for the first time in seven months, I had plenty of time to appreciate it.
