When people ask what it costs to publish with AI, they usually mean the model. In my pipelines the model is not the biggest line, and it is rarely where savings come from.
I run eight content pipelines that research, write, check and publish articles every day. Every call they make is logged with its cost, so I can say what a post costs rather than guess. Today a typical article costs around two US cents in AI and data calls. Getting there took a few months of measuring.
Where the money goes
A published post goes through several paid steps:
- Keyword and search data. Search volume, difficulty and what already ranks, from DataForSEO. This decides what to write about.
- A brief. A cheap model call that sets the angle, structure and links.
- The draft. One longer call to write the article in the site's voice.
- Quality checks. A smaller, fast model reviews the draft in batches, which costs about half the normal rate.
The surprise was that keyword data was the most expensive part, not the writing. Each post triggered several lookups, and many of them repeated work done the day before for a closely related topic.
What actually reduced the bill
In September 2026 I made three changes across all eight pipelines:
- Cache keyword lookups. If a pipeline asked about the same keyword recently, it reuses the answer instead of paying again.
- Reuse research across related articles. Articles in the same cluster share the research done for the first one.
- Only pay for data that changes a decision. Some lookups were being made "just in case". If a value never changed which topic got picked, I stopped buying it.
Together these cut the fleet's daily keyword-data spend from about $15 to $7.40, with no drop in what got published.
Plan for the day the money runs out
Prepaid API balances run out, and they tend to run out at night. In August 2026 two of my AI providers hit zero within the same hour, and every pipeline that depended on them stopped.
Since then each pipeline has a fallback provider, and I get a warning on my phone when a balance is getting low, well before it reaches zero. A pipeline that stops quietly for two days costs more in lost publishing than any of the API bills.
How to think about it for your business
- Measure cost per published post, not per month. A monthly bill tells you nothing about whether the system is efficient.
- Look at data costs before model costs. Cheaper models save less than you expect if the research step is wasteful.
- Budget for failure handling. Fallbacks and alerts are part of the cost of running automation properly.
If you are weighing up AI content for your own site and want a realistic cost estimate, send me a message.