Subscription or API
The same AI models can be paid for in two ways. With a subscription, you pay for people. With the API, you pay for what they consume.
Subscription: a price per person
A subscription gives one person access to a vendor’s application, for a fixed price every month. The seat is paid whether or not the person uses the tool. Usage stays within the vendor’s application.
- Fixed price per person and per month
- Used in the vendor’s application
- Usage limit included: beyond it, access is restricted or extended with paid credits
API: a price by consumption
The API gives access to the models themselves, from your own tools. Each call is billed by the volume processed, whatever the number of people. The organisation pays for what it consumes.
- Price per token, depending on the model
- Used in your own tools, including SmartAGT
- One meter for the whole organisation
- A cap you set yourself
The token, the billing unit
A token is a fragment of a word. In English, a hundred words are about 130 tokens.
The API counts tokens in both directions:
- Input, everything the model reads:
- the agent’s instructions
- the question
- the conversation history
- document excerpts
- Output, the answer, including the reasoning of models that think before answering
Output tokens usually cost more than input tokens, and a large model’s tokens more than a small one’s. Each new question sends the history back to the model: the longer the conversation, the more each question costs.
What each profile costs
| Profile | Subscription | API |
|---|---|---|
| Inactive | A full seat | Nothing |
| Occasional | A full seat | A few tokens |
| Intensive | A seat, within the usage limit | Many tokens |
In an organisation, usage varies from one person to another. The API matches the bill to each person’s actual usage, the subscription sets the same price for everyone.
Worked example, OpenAI public prices recorded in November 2026
A question to an agent that relies on documents: 4,000 input tokens, 1,000 output tokens. A month has 20 working days, with 5 questions a day for occasional use and 50 for intensive use. Prices in US dollars excluding tax.
| Model | One question | Occasional, per month | Intensive, per month |
|---|---|---|---|
| gpt-6-luna, economy | $0.0009 | $0.09 | $0.90 |
| gpt-6.1-sol, mid-range | $0.018 | $1.80 | $18 |
| gpt-6-astra, high-end | $0.09 | $9 | $90 |
As a subscription, ChatGPT Business costs $25 per person per month, or $20 with annual billing, whatever the usage and within the included limit.
Common misconceptions
“We already have subscriptions, let’s plug them into SmartAGT.”
A subscription is used in its vendor’s application. SmartAGT connects with an API key, which the vendor bills separately. Some plans include API credits, which are also consumed per token.
“The API is unpredictable.”
Cost follows usage, and SmartAGT makes it visible and lets you cap it:
- each call is valued at the offer’s price
- consumption is tracked for each account
- spending quotas limit it per day, per week or per month
“The API costs more.”
In the example, it costs less than a seat, except for intensive use of the high-end model. The choice of model matters as much as the volume.
Hosting your own models is a third way: you then pay for the servers rather than for tokens. See Inference.