A 1,500-Word Draft Costs a Tenth of a Cent. The Three Hours It Saves Are Yours to Bill

Why this pays

A client wants a 1,500-word article by Friday. Written by hand it is four hours; from a first draft in the client's voice it is one hour of editing. The draft itself costs between a tenth of a cent and ten cents, and the three hours saved stay with whoever wrote the brief.

The spread between the cheapest and dearest model on the page is under a dime an article. The cost that matters is your edit hour, and that is the number to quote from: a per-article fee priced on four hours, delivered in one, is the margin.

The play, in three lines

1. Turn one client's brief into a system instruction, with the reader, the length, the tone in a sentence, three things never to do, and three of the client's own paragraphs as examples. Write it once.
2. Run it in a browser (Gemini's AI Studio or Mistral Studio) or by one API request (OpenAI, Anthropic) with max_tokens above 2,000.
3. Cut the preamble, edit every fact against the client's sources, save the file, and time the edit. That time, not the token bill, is what you price.

The path above walks every step. What follows is the reasoning behind each and the numbers you quote from.

Where the money comes from

• The client, per article. They pay the same fee they paid when it took four hours. Nothing on the invoice says how the first draft was made; what they buy is the finished piece with every name and figure checked.
• The hour you keep. Three hours per article, or three more articles in the same day at the same fee.
• Not the model. The providers bill by the token: OpenAI from the first call with no plan fee, Anthropic on prepaid credits with no published minimum, Gemini and Mistral with a free tier at the door.

The pain you are selling the cure for

The client has a content calendar and no writer with time, and the last freelancer sent them something in the wrong voice that took two rounds of notes to fix. What they are paying for is the voice and the checking, not the typing. A brief that carries three of their own paragraphs as examples produces the voice on the first run; the edit pass produces the checking. The client never sees the draft and would not want to.

Step one: the brief is the system instruction

• Every provider puts the brief in the instruction layer, not the message. OpenAI's instructions parameter gives the model high-level instructions on tone, goals and examples of correct responses and takes priority over the prompt in input.
• Anthropic's best-practices page: be clear and direct, give the model a role, wrap 3 to 5 examples in <example> tags, put long material such as the client's previous articles near the top of the prompt above the instructions, and for source-heavy pieces ask the model to quote the relevant parts before it writes.
• The brief: who the reader is, the length in words, the section list, the tone in a sentence, three things never to do, three of the client's paragraphs. One block of text, reused every month.

Path A: in a browser, no code

• Google AI Studio opens the Playground with a new chat prompt. Run settings, top right, holds the System Instructions field; paste the brief, type the topic, click Run. Tighten the brief and run again until the register is right. Get code turns the prompt into API code; prompts can be saved and shared.
• Mistral Studio starts in Free mode with no credit card. Build, Prompts, New prompt saves the brief as a template with a title and description; a saved change creates a new version with the old one available for rollback; Share keeps it private or opens it to a workspace. For a client you write for monthly, that versioned prompt is the asset.
• The price of the free door: Gemini's pricing page prints, under every model, that Free tier content is used to improve Google's products and Paid tier content is not. Fine for your own topic; not for the client's strategy pasted in as an example.

Path B: one request

openai responses create --model "gpt-6-astra" --input "Write a one-sentence bedtime story about a unicorn." --raw-output --transform 'output.#(type=="message").content.0.text'

• That is OpenAI's own example. Swap the model for the one you priced, add --instructions with the brief, put the topic in --input. Pin a dated model snapshot in production so the voice does not shift between pieces.
• Anthropic: a POST to /v1/messages with model, max_tokens and a messages array; the system prompt holds the brief. The getting-started example sets max_tokens to 1000, and a 1,500-word draft is about 2,000 tokens, so leave it at 1000 and the article stops mid-sentence. The response reports input_tokens and output_tokens, so every call tells you what it cost.
• Mistral: /v1/chat/completions with mistral-large-latest. A 402 Payment Required means no payment method on the account, added under Admin, Subscriptions, Billing; a 429 means the Free mode rate limit, wait.

The size of the job, in tokens

• OpenAI's conversion: 1 token is approximately 4 characters or three-quarters of a word; 100 tokens are approximately 75 words, for English.
• 1,500 words out is about 2,000 tokens; a 300-word brief in is about 400. Those two numbers are the whole bill.

What the draft costs

The chart above is the output column as published, per million tokens: Together AI DeepSeek V4 Flash $0.28; OpenAI gpt-4o-mini $0.60, gpt-5.6-luna $1.20, gpt-5.6-sol $20.00 on promotional pricing through at least November 21, 2026; Mistral Large $1.5; Gemini 3.8 Flash $3.75 on the Paid tier until December 31, 2026 and $7.50 from January 1, 2027; Claude Haiku 4.5 $5, Sonnet 5 $10, Opus 5 $25, Fable 5.1 $50. Multiply by 2,000 tokens: a tenth of a cent on gpt-4o-mini, ten cents on Fable 5.1, with the 400 input tokens a fraction of that. The AI job cost calculator does it for any length; the Claude and OpenAI compare reads the two biggest lists side by side.

The edit pass

• The preamble. Anthropic's instruction, verbatim: Respond directly without preamble, do not start with phrases like Here is or Based on.
• The length. Claude Opus 5 defaults to longer responses than earlier models, and changing the effort setting does not reliably shorten them; ask for conciseness in the brief. On OpenAI, the snapshot pin is the fix for drift.
• Then yours, which no provider does: every number, name and claim against the client's sources, the section order against the brief, the register against the three example paragraphs. Save the file the client opens and send it.

Thirty drafts instead of one

• Anthropic's Message Batches API charges 50% of the standard price; most batches finish in less than 1 hour, anything unfinished expires at 24 hours and is not billed, a batch holds up to 100,000 requests or 256 MB, results stay downloadable for 29 days.
• Mistral's pricing page says batch processing halves its price as well.
• The same thirty drafts, half the bill, ready in the morning.

Who owns the draft

• OpenAI assigns to the customer all of its right, title and interest in the Output; the customer keeps ownership of its Input.
• Anthropic: the customer owns the Outputs and keeps all rights to Inputs; Anthropic assigns its interest in Outputs and may not train models on Customer Content.
• Mistral's open-weight models are Apache 2.0 for research and individual use; commercial deployments need a Mistral licence. Gemini: Paid does not train, Free does.

What is deleted when you stop

• Anthropic auto-deletes API inputs and outputs within 30 days. OpenAI deletes all Customer Content within thirty days of the agreement terminating and hands nothing back.
• OpenAI may limit or suspend access for a breach, without prior notice where necessary; Anthropic may suspend all access over a suspected Usage Policy breach and disclaims liability for the loss that causes.
• None of this touches a draft you already sent. All of it touches the one you left in the console.

What to do today

1. Take one piece you have already written for a client. Turn its brief into a system instruction with three of the client's paragraphs as examples.
2. Run it once in AI Studio or Mistral Studio, once by API with max_tokens above 2,000.
3. Time your edit pass on each. The token bill will be a few cents; the edit time is your real cost, and the number to quote from.