A ChatGPT subscription and the API run the same models, and the API costs three to five times more for the same 100 tasks, because OpenAI subsidises the $20 plan and does not subsidise the API. Those 100 tasks land somewhere between $60 and $100 on the API, which is $0.60 to $1.00 a task, and the per-task figure is the one you build a monthly budget from.
Is the API the same model as ChatGPT?
Yes. The big model labs make money in two places. The first is the subscription: you log into ChatGPT, Codex or Claude and use them on the lab’s own platform. The second is the business-to-business side, where other companies call the same models through an API and pay for what they use. Build an AI agent or an AI product and you are on the second one, paying for API credits.
The model underneath is identical. The chatbot is OpenAI’s product wrapped around the model. The API is the model with nothing wrapped around it.
Why does the API cost more than my $20 plan?
Because the subscription is subsidised and the API is not. OpenAI eats the difference on the $20 plan, and on the API you eat it. Run 100 tasks in ChatGPT and you might hit the ceiling of your $20 plan; run the same 100 tasks through the API, same output, same experience, and you can end up at $60, $80, sometimes $100. The work is identical, the model is identical, and the only thing that moved is which party covers the compute.
| Cost dimension | Subscription | API usage |
|---|---|---|
| What you pay | Flat monthly fee | Per unit of work |
| Who it suits | A person at a keyboard | An agent running on its own |
| The price | Subsidised by the lab | Unsubsidised |
| Cost of 100 tasks | Inside the $20 plan | $60 to $100 |
How much does an AI agent cost a month?
Roughly 90% of the time, an AI agent runs on API keys rather than on somebody’s subscription, so an AI employee sits in the right-hand column by default. Budget three to five times what the same outcome would cost you inside a chatbot.
The subscription price assumes a human at the keyboard. For most of the work you would hand to an agent there is no human available to sit there. An HVAC dispatcher is not typing every after-hours enquiry into a chat window at 2am, one at a time, until the phones go quiet. What the API bill buys is the 2am coverage, so the comparison that decides whether the agent is worth it is against the cost of a person covering that shift, not against your $20 plan.
How do I budget for an AI agent?
Price one job, per task and per month, before the agent is live. The only figure you need from outside your own business is the per-task range above.
- Count the runs. Take one job, say qualifying inbound solar enquiries, and count how often it happens. Suppose 200 enquiries a month.
- Multiply by the per-task range. $60 to $100 per 100 tasks is $0.60 to $1.00 a run, so 200 enquiries is $120 to $200 a month in API usage.
- Put it against the person. $120 to $200 versus what it costs to have a dispatcher qualify those same 200 enquiries, including the after-hours ones.
That range moves once the agent is live, and moving it down is a large part of the implementation work. A first version is usually expensive for avoidable reasons: it re-reads context it already has, or it runs on a model stronger than the job needs. Tuning is trimming those out without losing the output you want, and cheap and good pull against each other while you do it.
So the number to ask a vendor for is a monthly range per task. A per-token rate tells you nothing until you know how many tasks a month you are running, and turning one into the other means knowing what a token is and roughly how many of them one run of your job burns.
