The previous four lessons make the API look unanswerable. It is not. A subscription buys you an interface, a memory, file handling, image generation, mobile apps, and — the underrated one — a fixed number you can put in a budget. This lesson is where we stop selling and start counting.
The break-even, in one calculation
A typical assistant turn is about 500 input tokens (your question plus some conversation history) and 300 output tokens. On GPT-5 at $1.25 input / $10.00 output per 1M:
- Input: 500 ÷ 1,000,000 × $1.25 = $0.000625
- Output: 300 ÷ 1,000,000 × $10.00 = $0.003
- Per turn: $0.003625
- $20 ÷ $0.003625 = about 5,500 turns a month, or roughly 180 a day
So where does the money actually leak?
Three places, none of which is the price of a single seat:
- Seats nobody uses. The subscription that renews for someone who left, or who tried it for a week in March. This is the largest and most boring source of waste on this entire site.
- Bulk work on a conversational seat. Classifying a thousand tickets by hand in a chat window, when lesson 3 does it for a fraction of a cent. This is where the API wins by 100×, not on conversation.
- Frontier defaults on trivial work. Sending every request to the most capable model because it is the one in the dropdown. Lesson 4 is the fix.
| Your usage | Cheaper option | Why |
|---|---|---|
| Conversational, under ~150 messages/day | Subscription | You will not reach the break-even, and you get the app, memory and files |
| Bulk, repetitive, same prompt shape | API | No interface needed, and a cheap model almost certainly passes |
| Both, which is most teams | One seat + API for the bulk | Stop doing batch work by hand in a chat window |
| You cannot say which | Measure for two weeks | Any decision made now is a guess with a price tag |
Knowledge check
You send about 40 chat messages a day and also classify 5,000 records a month. What is cheapest?