Speculation about the cost of AI-assisted publishing is cheap. Here are our actual numbers.

Across four generated bulletins covering thirty-nine stories, the script writer consumed 47,924 tokens. At the rates for the small model we use for this, that is approximately $0.0093 total — roughly $0.0023 per bulletin. Generating a fifteen-story bulletin every hour, around the clock, would come to something in the region of three dollars a month.

Why it is so cheap

Because the model is doing the least interesting part of the job. It is not researching, not deciding what matters, not writing from scratch. Story selection is a scoring function over articles we already hold — recency, engagement, editorial flags, geography, completeness. The script is largely extractive. The model smooths transitions and writes a wrapper. That is a small, bounded task, and small bounded tasks are what the cheap models are good at.

Where the real cost sits

Not in tokens. Text-to-speech and avatar generation are one to three orders of magnitude more expensive per minute of output than the script that produced them, which is why our default configuration synthesises narration in the reader's own browser at no cost at all. Video generation per page view would be ruinous, which is why bulletins are generated once and served as static output rather than rendered on demand.

The implication worth sitting with

If assembling a competent-sounding news bulletin costs a fifth of a cent, then the constraint on flooding the web with synthetic news copy is not economic. It never will be again. Whatever keeps a publisher honest from here has to come from editorial process — sourcing, validation, correction, willingness to say "we do not know" — because cost is no longer doing any of that work.

That is the reasoning behind the validation gate described in What an AI News Anchor Can and Cannot Do.