Every CEREBRO run meters itself. The number above is the sum of the cost_usd field in the frontmatter of all 87 published briefings, in integer hundredths of a cent, summed once and divided once, so it is exact rather than approximately right.

What it is: the Claude API spend of the triage and digest passes, plus the Hacker News comment enrichment. Across the corpus that is 365 model calls and 22,496,242 tokens, of which 9,052,158 were cache reads (the cheap kind), and that is why the figure is what it is rather than several times larger.

What it is not: it excludes the machine it runs on, the bandwidth, the scrapers, the site’s own hosting, and every hour spent building the thing. It is the model bill, not the cost of the project. Anyone quoting it as “what an AI news digest costs” is quoting the smallest line of the invoice.

It works out at $0.5825 a briefing and $0.0259 per signal that actually made it through, on 258,577 tokens a day. The colophon explains which stages spend that and why the pre-rank exists.