What an AI-Written Article Costs, Priced Token by Token

Seven drafts of a 1,500-word article, run on Claude Fable 5.1 at the top of Anthropic's published table, cost $1.88 in tokens. On a mid-priced model most people would actually use for this, the same seven drafts cost $0.38.

Those two figures are the answer. The rest of this page is the arithmetic behind them, written out so you can redo it with your own article length, your own prompt and your own honest count of how many times you re-ran it.

Turn the job into tokens first

A rate per million tokens tells you nothing until you know how many tokens your job is. Anthropic publishes what you need to work that out. OpenAI publishes prices per token and no way to turn an article into tokens, which is a gap worth naming now, before it quietly gets filled with a guess.

For Claude, a token approximately represents 3.5 English characters. Characters, not words, so there is one step in between. Two finished articles on this site measure 4,977 characters across 894 words, and 5,158 characters across 936 words. Call it 5.6 characters a word, the higher of the two, because rounding in the direction that flatters the estimate is how estimates go wrong.

Anthropic's own pricing page answers the same question with a different number, calling 1 token approximately 4 characters or 0.75 words in English. The two pages disagree by about 14%, and this one uses 3.5, because it produces the larger token count and so the larger bill.

That gives two numbers to work with:

• A 1,500-word article is about 8,400 characters, so about 2,400 tokens of output.
• A real prompt is larger than people expect. A 1,000-word brief plus two reference documents of 1,500 words each is 4,000 words, about 22,400 characters, about 6,400 tokens of input.

So one pass is 6,400 tokens in and 2,400 tokens out. That is the count for the older tokenizer, and most of the table below does not use it.

The 30% nobody carries through

Anthropic's pricing page carries a note that gets skipped. Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text, while Claude Sonnet 4.6 and earlier models use the previous one.

Sonnet 5, Opus 5 and Fable 5.1 are all past 4.7, so all three are on the new tokenizer. Haiku 4.5 is not. That means the same article is two different sizes depending on which model reads it:

• On Haiku 4.5, one pass is 6,400 tokens in and 2,400 out.
• On Sonnet 5, Opus 5 and Fable 5.1, the same pass is about 8,320 in and 3,120 out.

Nothing about the article changed. The meter did. A 30% rise in the count is a 30% rise in the bill, it applies to every model shipping from here on, and it arrives without any change to the rate card, which is exactly why it goes missing from the comparison posts.

What one pass costs

Every figure in the first two columns is read off Anthropic's own pricing page. The last column is those rates against the token counts beside them.

gpt-4o-mini and gpt-5 are on this page too, and they are deliberately not in this table. Their rate cards are published, at $0.15 and $0.60 per million and at $1.25 and $10.00 per million. Their tokenizer is not: the pricing page those rates come from states no character-per-token ratio at all. Running the Claude ratio over an OpenAI model would produce a number that looks exactly as precise as the rest of the column and rests on nothing, so the cost per article for those two models is not on this page. If you need it, count your own tokens with the provider's tooling and use the rates above.

Look at where the money sits. Output is priced at five times input on every Claude model in the table, and between four and eight times on the two OpenAI models, and your article is the output. Choosing a model on its input rate, which is the number the marketing tends to lead with, optimises the cheap half of your bill.

The multiplier nobody puts in the estimate

Here is where every published estimate goes wrong, including the friendly ones. They price one pass.

Nobody ships the first draft. A realistic run is a first attempt, a structural rewrite, two rounds of "this section is thin", a fact-checking pass, a tightening pass, and a final read. Seven passes, and seven is not a pessimistic number for something you intend to put your name on.

The passes are not equal. Every pass after the first sends the brief again, plus the draft it is fixing, plus what you want changed. On the older tokenizer that is 6,400 plus 2,400 plus 200, so about 9,000 tokens of input each time. On the newer one it is 8,320 plus 3,120 plus 260, about 11,700.

Seven passes on Haiku 4.5 therefore cost 6,400 plus six times 9,000, which is 60,400 tokens in, and seven times 2,400, which is 16,800 tokens out. On the three newer models it is 8,320 plus six times 11,700, which is 78,520 tokens in, and seven times 3,120, which is 21,840 out.

Run those through the rate cards and one 1,500-word article costs $0.14 on Haiku 4.5, $0.38 on Sonnet 5, $0.94 on Opus 5 and $1.88 on Fable 5.1.

Note the ratio. Seven passes cost about eight times one pass, not seven, because the rewrites pay twice: once to read the draft and once to write the new one. If you quote yourself the single-pass number, you are out by a factor of eight before you start.

Against what the same words cost from a person

The Editorial Freelancers Association publishes a rate chart. Work-for-hire blog posts run 25.0 cents to 45.0 cents a word, which is $375 to $675 for 1,500 words. Ghostwriting for articles and essays runs 15.0 cents to 30.0 cents a word, so $225 to $450 for the same length.

Set that beside $0.94. You could run the seven-pass pipeline on Opus 5 about 240 times for the price of one article at the lower of those two rates. The chart carries cheaper writing rows than either, so treat 240 as the conservative end of the gap rather than the size of it.

That comparison is real and it is also a trap, because it is not the comparison you are actually making.

The number on this page that is actually large

The token bill is not what an AI-written article costs you. Your time is.

Seven drafts have to be read, judged and marked up by somebody, and that somebody is you. Ninety minutes is a fair estimate for reading seven versions of 1,500 words and knowing which parts to keep. That work is editing, not writing, and the same chart prices editing separately and lower: $45.00 to $60.00 an hour to proofread business and marketing copy, $50.00 to $60.00 to copyedit it, and $55.00 to $75.00 or $60.00 to $75.00 for line and developmental editing. Ninety minutes of it is worth $67.50 to $112.50.

Against $67.50 of your own time, the bottom of that range, a token bill of $0.38 is about half of one percent of what the article cost to produce.

This is the finding that changes decisions. Arguing about which model to use to save half a dollar an article is time spent on a rounding error. The two things that move the real number are how many passes you need before a draft is usable, and how long each pass takes you to review. A better prompt that gets you to a keeper in three passes instead of seven saves you an hour, which is worth several hundred times what it saves in tokens.

Where the meter does start to bite

Per-token pricing gets serious when the human review stops scaling with it, which means volume.

Two hundred articles a month, seven passes each, is $75.09 on Sonnet 5 and $187.72 on Opus 5. Those are real monthly numbers, and they are the point at which model choice earns its keep. Two levers exist before you drop to a weaker model. Anthropic offers 50% off for batch processing if you can wait for the answer, and OpenAI publishes batch rows at half the standard price, with gpt-5 at $0.625 input and $5.00 output. Halving $187.72 is worth more than any prompt tweak.

The other lever is not paying twice for the same words. Neither provider hands your drafts back later, so a draft you did not save is a draft you buy again.

Do this before you pick a model

Open the last article you produced this way. Count its characters and divide by 3.5, then add 30% if the model is Claude 4.7 or later, and you have your output tokens. Count the brief and everything you pasted in, do the same, and you have your input tokens for one pass. Then answer the honest question: how many times did you actually run it?

Multiply, and you will have a number under two dollars. Then time yourself on the next review round, and you will see where the article really costs you.