How Much Does an AI-Generated Article Cost? I Paid and Counted

👁 114 reads

When you ask “how much does an AI-generated article cost,” you usually get an answer calculated from a price list. I decided to pay and find out. Eight models, the same prompt, an article about nine hundred words long, cost read from the actual invoice, not estimated.

The cheapest wrote the text for $0.00039. The most expensive for $0.18913. That’s four hundred eighty-five times more for the same length. And one model didn’t write a single word but still sent an invoice.

Results

The entire test cost $0.43653.

The Model That Charged for Silence

GLM 5.3 Flash received a limit of eight thousand tokens for its response. It used all eight thousand on internal reasoning and left none for the actual text. Result: zero words, 82 seconds of waiting, and an invoice for $0.00402.

This isn’t a fluke or a one-time error. In this model you can’t disable reasoning, because the API rejects such attempts with an error. For short tasks, this just inflates your bill. For longer ones, it can consume your entire response.

If you’re building anything that runs unattended, this is exactly the kind of failure you won’t see in logs: success status, non-zero invoice, and an empty output file.

Cheapest Doesn’t Mean Worst, But You Need to Know What to Look For

The two cheapest models differ by less than a penny in cost, but there’s a chasm in quality.

DeepSeek V4 Flash wrote 922 words for $0.00039, without a single reasoning token, and hit the target length within two percent accuracy. The text we tested in Polish begins with a sentence that sounds human: “Kiedy po raz pierwszy usłyszałem o asystentach AI, pomyślałem, że to technologia zarezerwowana dla dużych korporacji” (When I first heard about AI assistants, I thought it was technology reserved for large corporations).

Ministral 14B was even slightly cheaper and the fastest in the entire lineup, but it opened the text with: “W dzisiejszym świecie, gdzie efektywność i automatyzacja są kluczowe” (In today’s world, where efficiency and automation are key). On top of that, it inserted twenty-nine em dashes, a punctuation mark that in Polish text immediately betrays machine generation, because almost no one uses it from the keyboard.

Both cost the same. One needs editing, the other needs rewriting.

Who Hit the Target Length

I asked for about nine hundred words. This is a test of obedience, not creativity, and it turned out more interesting than I expected.

• Hit within three percent accuracy: Kimi K3 (903), GPT-6 Astra (906), Claude Fable 5.1 (886), DeepSeek V4 Flash (922).
• Overshot: Gemini 3.8 Flash by fifteen percent (1,032), Ministral 14B by thirty-two percent (1,192).
• Fell short: Claude Haiku 4.5 wrote 596 words instead of nine hundred, thirty-three percent too few.

This has practical significance. If you’re ordering text for a specific space in a page layout, a model that regularly comes in thirty percent short means you’ll be adding content manually.

What This Costs at Monthly Scale

Assume three articles per day, or ninety per month.

Only in this table do you see what’s really at stake. For one text, the difference is invisible. For ninety, it’s the difference between four cents and seventeen dollars, for the same length.

Time Costs Too

Kimi K3 needed 182.7 seconds, over three minutes for one article. GPT-6 Astra did the same in 51 seconds, and Claude Haiku 4.5 in 16, though shorter.

For manual work, three minutes is nothing. For ninety texts, that’s four and a half hours of pure waiting per month, versus seventy-six minutes with Astra.

Where These Invoice Differences Come From

The answer lies in reasoning tokens - text the model writes to itself and that you pay for at the output rate.

• DeepSeek V4 Flash, Ministral 14B, and Claude Haiku 4.5: zero reasoning tokens.
• GPT-6 Astra: 176. Claude Fable 5.1: 715.
• Gemini 3.8 Flash: 985.
• Kimi K3: 4,667, for an article of 903 words.
• GLM 5.3 Flash: 8,000, the entire available limit and nothing beyond it.

There’s a second thing few people remember when calculating: the price on the model card may not be the price you pay, because the same model is hosted by dozens of companies with their own rates, and the spread reaches fourteen times.

What This Means in Practice

For bulk text processing, use DeepSeek V4 Flash. Zero reasoning, accurate length, invoice at rounding-error level. Just check the output, because cheap doesn’t mean flawless.

For text going to publication without revisions, pay more. GPT-6 Astra wrote the only text that starts with something concrete rather than a warm-up: “Nie zaczynam od zakupu narzędzia ani od próby zautomatyzowania całej firmy“ (I don’t start with buying a tool or trying to automate the entire company). That’s the difference you pay ten dollars a month for instead of four cents.

Always check three things in the output: em dashes, openings like “in today’s world,” and whether the model wrote anything at all. All three can be checked automatically and all three happened in this test.

And most importantly: cost is not the same as value. This test measures what you pay for nine hundred words in Polish - a useful proxy for non-English, heavily inflected languages - not whether those nine hundred words are useful to anyone. For how I check language quality itself, I wrote about it in a test of ten models on grammar tasks, and instruction-following in a test of eleven models on code and commands.

How I Measured This

One prompt, sent September 10, 2026, identical for all eight models, with an eight-thousand-token limit on responses. The cost of each query was read from the billing field returned by the API. Word count, diacritics, em dashes, and headings were counted programmatically on the raw response.

One run per model is too few to talk about averages, and I’m not claiming that. I’m claiming this is how it came out on the first attempt, and that a four-hundred-eighty-five-times difference in price is too large to explain by variance.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Why This Couldn’t Be Avoided

The model that burned through the entire limit on reasoning doesn’t let you disable it, and it’s not alone. I reviewed eighteen models: seven reject such attempts with an error, including GPT-6 Astra, Claude Fable 5.1, and both Gemini Flash models.

Kinetyra