When a $20 seat beats the API for agent tasks, and when it does not
A $20 seat with an agent mode is a fixed price for a capped amount of looping. The API is a per-step meter with no cap unless you set one. Work out which is cheaper at your volume using the reference trace.
In this lesson you will
→Convert a $20 seat into a number of agent runs on each model
→Name what a seat gives you that the API does not, and the reverse
→Pick seat, API, or both for your own agent usage
Lesson 2 priced a 12-step run at $0.0089 to $0.47 depending on the model. A $20 seat (ChatGPT Plus, Claude Pro, Google AI Pro at $19.99) sells you agent features for a flat fee with a usage limit that resets over time. The question is not which is better. It is which is cheaper for the number of runs you will actually do.
Figure 1.A subscription is a flat line: the same price whether you run one agent task or a hundred, until the cap. The API is a slope: cheap at low volume, uncapped at high volume unless you cap it.
How many runs is $20?
$20 divided by the cost of one 12-step reference run. Uncached; caching roughly triples the Sonnet 5 and GPT-5.4 figures.
Model
12-step run
Runs per $20
Per working day (22)
DeepSeek V4 Flash
$0.0089
about 2,240
about 100
Gemini 3.8 Flash
$0.139
about 144
about 6
Claude Sonnet 5
$0.370
about 54
about 2
GPT-5.4
$0.471
about 42
about 2
Read the last column. If you run two or three agent tasks a day on a frontier model, the API and the seat cost about the same, and the seat comes with an app, a browser tool, file handling and no setup. If you run twenty a day, the API on a frontier model costs $200 a month and the seat's cap will stop you long before that anyway. If you run two hundred a day, you are not a seat customer; you are a routing problem, and lesson 2's cheap models are the answer.
What each option gives you that the other does not.
$20 seat with agent mode
API
Price shape
Flat, capped; the cap changes and is not published as a fixed number, so check the vendor page
Per token, uncapped unless you set caps (lesson 5)
Model choice
Whatever the vendor routes you to
Any model, so the cheap ones from lesson 2
Setup
None
A key, a script or a tool that takes a key
Runs on a schedule while you sleep
Generally no
Yes, which is the point and the danger
Tools it can use
The vendor's: browser, files, code
Whatever you connect
Cost visibility
None per task
Exact, per call
Knowledge check
You run about 3 agent tasks a working day, interactively, and want the vendor's browser tool. Which is cheapest?
Lesson FAQ
▸Is ChatGPT Plus worth it for agent tasks?
For a few interactive agent runs a day, yes: the same runs on the GPT-5.4 API cost about as much and you would have to build the tools yourself. It stops being worth it when you need runs on a schedule, high volume, or a cheaper model, which are all API territory.
▸How many agent runs does $20 buy on the API?
On the site's 12-step reference trace, about 42 on GPT-5.4, 54 on Claude Sonnet 5, 144 on Gemini 3.8 Flash and roughly 2,240 on DeepSeek V4 Flash, uncached. With prompt caching the frontier figures roughly triple.
▸Should I buy Claude Max or ChatGPT Pro for more agent usage?
Only if you would use the extra runs on a frontier model anyway. Price the same volume on the API with a cheaper model first; $100 buys about 720 Gemini 3.8 Flash runs, which is more agent work than most people do in a month.
▸Can I use both a seat and the API?
Yes, and it is the common answer: one $20 seat for interactive tasks and the app, plus API spend on a cheap model for anything repeated or scheduled. The two bills stay small for different reasons.
Finished reading?
Mark it done to track your progress through the course.
If our calculators helped you cut down on hidden AI wallet leaks, thanks for using them. A tiny fraction of your savings is what keeps our pricing indexes updated daily.
Not sure which AI is cheapest for your use case? Find out in 30 seconds — no signup required.