Price your AI product by the job, not the token
In one week, AI model prices moved in both directions.
On October 7, Anthropic released Claude Haiku 5.5 and said it costs around 75% less to run than Haiku 4.5 on average. On October 13, Google launched Gemini 4 Argon at an introductory $2 per million input tokens and $10 per million output tokens. After November 30, that price doubles to $4 and $20. On October 14, OpenAI told developers that dated GPT-4o, GPT-4o mini, GPT-4.1, and GPT-4.1 mini snapshots leave its API on January 14, 2027.
Cheaper. More expensive. Gone.
If your price is tied to tokens, your margin is tied to someone else’s release calendar. Your customer should never feel that. Price the job your product finishes, and treat the model as a cost you manage behind the scenes.
Your customer is not buying tokens
A practice manager does not want “50,000 tokens of AI.” She wants 40 insurance pre-authorizations drafted before Monday. A procurement lead wants every vendor contract checked against the company’s standard terms.
That is the unit of value. One finished, checked piece of work.
When you charge per message, per credit, or per token, you hand the customer a meter they cannot read. They start rationing usage. They worry about surprise bills. Worse, you teach them to compare you with a raw model subscription, which is a comparison you lose.
When you charge per job, the conversation changes:
- Per pre-authorization drafted, not per prompt.
- Per contract reviewed, not per page processed.
- Per weekly report delivered, not per seat.
Your domain experience is what makes this possible. You know what a “done” job looks like in your industry, how long it takes a person today, and what a mistake costs. A generic tool cannot price that way because it does not know where the job ends.
Model prices will keep moving, so build a cost buffer
Look at what one job actually costs. Say your product reviews one contract using 60,000 input tokens and 8,000 output tokens.
- On Gemini 4 Argon at the introductory rate, that is about $0.20 per review.
- On December 1, the same review costs about $0.40.
- On Claude Haiku 5.5, at $0.10 and $0.50 per million tokens for prompts under 100,000 tokens, it is about $0.01.
Token counts differ by model and tokenizer, so treat these as rough examples, not quotes. The point is the spread. The same job can cost 40 times more depending on the model, and the price of one model can double on a date you did not choose.
If you charge $15 per review, none of these numbers threaten your business. If you charge $0.50 per review because you priced “a little above cost,” a single pricing change can wipe out your margin.
Three habits protect you:
- Know your cost per job, not per call. Count retries, long documents, failed runs, and the human review step. The model bill is often the smallest line.
- Set a margin floor. Decide the lowest acceptable gap between price and cost per job, then watch it weekly.
- Leave room for the expensive day. Price so that your most costly realistic model setup still leaves you profitable.
Model retirements are a pricing event, too
OpenAI’s deprecation notice gave developers about three months before the GPT-4o and GPT-4.1 snapshots shut down. The New Stack noted that OpenAI’s own policy promises at least six months for generally available models. Fine-tuned versions of those models shut down on the same day.
The recommended replacement, GPT-6 Luna, lists at a lower token price. But a migration is never just a price swap. Outputs change. Tool calls behave differently. Your quality checks need to run again before customers see the new results.
That work costs time, and time is part of your cost per job. If your pricing assumes nothing ever changes underneath you, every forced migration comes out of your margin.
Build two things now:
- A short list of acceptable models for each job. If one is retired or repriced, you already know the fallback.
- A small test set of real jobs. Run it against any new model before you switch, so a cheaper model does not quietly deliver worse work.
Price the outcome, then pick the model
Most first-time founders work backward from cost: “The model costs X, so I’ll charge X plus something.” Flip it.
- Start with the customer’s current cost. How many hours does this job take a person today? What happens when it is late or wrong?
- Set a price that is clearly worth it. A fraction of the current cost is an easy yes for a buyer.
- Then choose the cheapest setup that meets your quality bar. Haiku 5.5 is the first Haiku model with an adjustable effort setting, which lets you spend more thinking on hard cases and less on routine ones.
- Review the cost per job monthly. When model prices fall, you keep the margin or reinvest it in better quality. When they rise, your buffer absorbs it.
This is also a stronger fundraising story. Investors hear “we charge per completed job, our cost per job is tracked weekly, and we’ve already switched models once without changing customer pricing.” That sounds like a business. “We mark up tokens” sounds like a reseller.
The takeaway for domain-expert founders
The model market will keep shifting. Prices will drop, rise, and reset with each launch. Models will retire on someone else’s timeline.
Your advantage is not access to a model. Everyone has that. Your advantage is knowing exactly which job matters in your industry, what “done” looks like, and what it is worth to the person paying.
Price that. Let the model be a cost you manage, not the thing you sell.
AI Product Accelerator is a 12-week, mentor-led program for turning what you already know into an AI product you build, launch, and own, including how to package and price it. Explore the program if you want help defining the job your product finishes and the price a customer will pay for it.