A short piece of fiction, written by API Evangelist. The reader is imagined; the post, and every number and quotation from it, is real: "OpenAI API Pricing: What Every Model Costs Per Million Tokens," Kin Lane, November 6, 2026, https://apievangelist.com/2026/11/06/openai-api-pricing-what-every-model-costs-per-million-tokens/
Marisol ran support operations for a company that sold scheduling software, and on Friday afternoon her manager asked her what it would cost to have a model sort the ticket backlog every night. She said she'd find out. Then she searched, and the first thing she read was that nobody can give her one number.
She found that annoying. She had wanted a figure to paste into a message. Instead the post, which Kin Lane wrote on November 6th, 2026, told her the price depends on which model she calls, whether the tokens go in or come out, and how fast she needs the answer. She read the table anyway. It listed twelve models, and she noticed it was twelve of twenty-seven. That bothered her a little. She was being shown a slice, and the full list lived on the author's own site, APIs.io. She didn't think that was dishonest, but she noted it. Whoever picks the slice has shaped the question.
Still, some things stuck. Output tokens cost between four and eight times as much as input, so a job that reads long tickets and writes back one label should be cheap on the expensive side. She had assumed the opposite, that the writing was the light part. And the spread between models was bigger than she'd have guessed. The post said that choosing the model is the biggest pricing decision you will make, and she believed it, because the gap between the top and bottom of the table made her own preference for the newest model look like a habit, not a decision.
The part she lingered on was Batch. The post used her exact job as its example: classifying a backlog of support tickets overnight, at half the standard rate. Nobody at her company was waiting on those labels at two in the morning. She felt faintly foolish for having pictured the fast, pricey version of the job. The post put it plainly: If your users are not waiting on the response, Batch is the easiest saving available.
What she did not enjoy was the last section. Prices move often, the author admitted, and the previous record had listed three models that were gone from the page. So check the date on any price you read, including mine. She appreciated the honesty, and it also meant she could not hand her manager any figure from this page and call it settled. The post had been read from OpenAI's pricing page on November 5. By the time she finished her budget it might be stale.
So she kept her ambitions small. She opened a spreadsheet and made four columns: a ticket's tokens in, tokens out, the model, and the date she read the price. She pulled one ordinary ticket from the queue and its typical label, counted the tokens, and filled in the first row. Then she wrote a note to herself for Monday: price it on two or three models, not just the shiny one, and open OpenAI's own page before sending anything to her manager.