Research and sourcing
Finding, reading and citing primary sources. A model can produce citations; verifying they exist and say what the draft claims is human time.
A 1,500-word article costs $0.0164 in API fees on a mid-range model and $225 from a freelance writer at $0.15 a word — roughly 13,736× more. That gap is real, but it compares a raw draft with finished copy, and the two are not the same product.
A 1,500-word article, against $0.0164 on GPT-4.1.
| Writer | Per word | Per article | vs API |
|---|---|---|---|
| Content mill / bulk marketplace | $0.03 | $45.00 | 2,747× |
| Competent generalist freelancer | $0.15 | $225 | 13,736× |
| Experienced specialist | $0.40 | $600 | 36,630× |
| Subject-matter expert | $1.00 | $1,500 | 91,575× |
Per-word rates are typical market bands, not quotes — enter your own on the calculator. The API figure is generation only.
The two columns above are not the same product, and treating them as one is how content budgets go wrong.
Finding, reading and citing primary sources. A model can produce citations; verifying they exist and say what the draft claims is human time.
The single largest hidden cost. Confident wrong sentences take longer to catch than obviously wrong ones, and they carry real liability in regulated topics.
Structural editing, cutting, and making a draft sound like your publication rather than like a model. Repeatable, but not free.
A freelancer carries professional responsibility for what they file. An API carries none, so the risk stays with you.
At a $50/hour editorial rate, thirty minutes of cleanup costs the same as generating roughly 1,526 drafts on GPT-4.1. This is the whole argument in one line: once a human touches the draft, the token cost stops mattering. A model that needs an extra fifteen minutes of repair per article is more expensive than one costing ten times the tokens, at any realistic volume.
Which is why the useful question is not "is AI cheaper than a writer" — it obviously is, per word — but "which model produces a draft my editor will accept fastest".
GPT-4.1 against a $0.15/word freelancer, generation cost only.
| Articles | API | Freelance | Difference |
|---|---|---|---|
| 10 | $0.1638 | $2,250 | $2,250 |
| 50 | $0.8190 | $11,250 | $11,249 |
| 100 | $1.64 | $22,500 | $22,498 |
| 500 | $8.19 | $112,500 | $112,492 |
The difference column is the budget you have left for editing — not money saved.
Original reporting, interviews, opinion with a byline that matters, regulated topics, and anything where being wrong is expensive.
First drafts from a clear brief, format conversion, volume work on well-understood topics, and outlines an editor will rework anyway.
The common answer: a model drafts, a human edits and verifies. Cheaper than pure human, safer than pure model.
If nobody will read it, the cheapest article is the one you do not publish. Volume without demand costs more than it returns — and so does a page the crawlers that send readers cannot reach. Check which search and AI bots can crawl it before you scale up.
Generation is now close to free — $0.00080 an article on GPT-5 nano. Everything that made a good article expensive is still expensive.
Run your own numbers →