$6 spent in 2 days before I turned it off. On something that was to be free.
@Apeksha Ghosh and I started on https://happinews.blog a few months ago, a personal project to cut through the noise and find positive news. 40+ editions later, it is still running.
I built it using Claude Code. Mainly to learn its quirks, but also to know how to talk to developers using AI assistants.
I created it to be mainly a non-AI app. The only place AI was brought in was for summarizing items, using Gemini 2.5 Flash Lite, a model offered by Google at no charge within a daily usage limit.
2 days of use got the bill to $6. For every item that was summarized, Google was charging me, because my Google Cloud account had moved to the paid tier some time back. On a paid account, the free model is not free. Every call is billed from day one.
By day 3, I had turned it off. Summarization is now triggered manually, only for already shortlisted items. The total cost over the next 50 days was under $0.10.
There was nothing wrong with the technology. The assumption underneath had simply never been correct.
I read about enterprise AI programmes that exhaust their annual budgets within the first quarter. The technology usually works. The failure is usually a pricing assumption made during planning that wasn't verified at go-live. At that scale, $6 becomes $600k. The process that has just started relying on it needs to halt to fix the billing assumptions.
AI costs aren't stable. Terms have been changing, tiers have been shifting. When was the last time you checked your AI pricing assumptions?
This is the first of a few posts about what building and running happinews has taught me about using AI in practice.