The video I posted last week cost about $52 in tokens at API prices (it came out of my subscription, not an extra bill). The surprise was where the money went.
I ran everything in Claude Code on the cloud, so I wouldn't have to install anything or tie up my PC. Without a graphics card, each full render took about 2 hours. Rendering doesn't use tokens, the AI just waits.
The cost came from coming back. After an hour idle the cached context expires, so every time I returned from a render, the AI had to load the whole conversation again: code, logs, review images. About $33 of the $52, almost two thirds, went to the AI catching up on what it already knew.
The funny part is that I know this, and that's why I am very strict about keeping context lean in my day-to-day work with Claude. But video was the first time it got away from me, because the work had long waits naturally built into it.
Next time I'll split it by phase, one session for the brief and storyboard, one for building, one for rendering and review, with everything saved in between so each session starts small.