Choosing between an API and a subscription.

Goal

Decide whether one workflow should run on per-seat AI subscriptions or through an API, with the costs and trade-offs written down.

Who it is for

Owners, finance managers and operations leads in New Zealand businesses, and anyone asked “should we just build this on the API?”

Time

30 minutes.

You need

ChatGPT, Claude, Copilot or Gemini, the fictional workflow and price table below, and a calculator.

You will produce

A side-by-side monthly cost for one workflow and a recommendation with its reasons.

What is the difference between a subscription and an API?

A subscription is a monthly fee per person for a finished app: ChatGPT, Claude, Microsoft Copilot or Gemini, used in a browser, desktop app or inside Microsoft 365 or Google Workspace. An API is metered access to the same family of models, called by your own software or an automation tool, and billed per token of text in and out. There is no app; someone has to build the part people see.

They are separate products. OpenAI’s ChatGPT Business overview states that the subscription “does not include API usage”, which is billed separately.

How do the costs compare?

Subscriptions are predictable: seats times price. As listed on 19 October 2026, ChatGPT Business showed New Zealand visitors NZ$33 per user per month billed annually for a Standard seat, and Claude Team showed $40 per Standard seat per month billed annually, with prices stated to include 15% GST.

API costs scale with volume and model choice. Prices are quoted in US dollars per million tokens (MTok), split between input and output. A token is a chunk of text; Google Cloud’s pricing documentation uses roughly four characters per token as an approximation, which is close enough for planning.

Gemini 4 Argon rises to US$4 input and US$20 output from 1 December 2026. Re-check every price before you budget.

Exercise: cost one workflow both ways

The business is fictional; the prices are from the table above.

Illustrative workflow. Fictional tour operator, Rotorua.

Task: draft replies to booking enquiry emails for staff to check
Volume: 1,500 enquiries a month
Staff handling enquiries: 3
Per enquiry sent to the model:
  instructions and FAQ text   about 6,000 characters
  the customer's email        about 2,400 characters
  draft reply returned        about 1,200 characters
Systems: Gmail inbox, booking system with no AI features

1. Run the costing prompt in the worked example below with the workflow and price table.
2. Check the arithmetic. At four characters per token, each enquiry is about 2,100 input and 300 output tokens. Confirm the monthly totals of 3.15 MTok input and 0.45 MTok output.
3. Cost the subscription route. Three seats at the listed price, monthly and annually.
4. List the hidden API costs. Ask the tool what else the API route needs. You should see building the inbox connection, a review screen, logging, testing and maintenance.
5. Estimate the build in hours, not dollars, then price those hours at whatever your developer or provider charges.
6. Decide with the rule below and write two sentences explaining why.

Worked example: the costing prompt

Using the workflow and the price table below, calculate:
1. Input and output tokens per enquiry, at 4 characters per token.
2. Monthly input and output tokens (in millions).
3. Monthly API model cost in USD for each model in the table.
4. Monthly cost of 3 subscription seats at NZ$33 per seat.
Show every step. Then list the costs of the API route that are
not model fees. Do not invent prices that are not given.

[paste workflow and price table]

Example output (illustrative and abridged; your tool’s wording will differ):

Per enquiry: (6,000 + 2,400) / 4 = 2,100 input tokens;
1,200 / 4 = 300 output tokens.
Monthly: 1,500 x 2,100 = 3.15M input; 1,500 x 300 = 0.45M output.

Claude Haiku 5.5 or GPT-6 Luna:
  3.15 x $0.10 + 0.45 x $0.50 = US$0.54 a month
Claude Sonnet 5.5, GPT-6.1 Sol or Gemini 4 Argon (intro):
  3.15 x $2.00 + 0.45 x $10.00 = US$10.80 a month
Subscriptions: 3 x NZ$33 = NZ$99 a month.

Non-model costs of the API route: Gmail integration, a screen for
staff to approve drafts, error handling, logging, testing on real
enquiries, monitoring and ongoing maintenance.

Model fees under US$11 a month look like an easy win, but the build and upkeep cost more than years of fees. The API route pays off when it removes copy-and-paste work at volume, not because tokens are cheap.

How do you decide?

Choose subscriptions when the work is varied, a person handles each item anyway, volumes are modest, and you want results this week. Choose the API when one task repeats at volume, it must run inside your own systems, you need structured output or logging, and someone will own the build. Many businesses run both: seats for everyday work, an API build for one high-volume process.

Data handling rarely decides it alone. OpenAI’s enterprise privacy page says Business, Enterprise and API data “isn’t used for training our models” by default, and Anthropic says the same for its commercial products. One real difference: none of OpenAI, Anthropic or Google offers AI processing inside New Zealand, but Amazon Bedrock’s Australia and New Zealand profile keeps Claude inference in that geography, an option only an API build can use.

Where this goes wrong

• Comparing token prices only. The cheapest model is irrelevant if nobody maintains the integration.
• Paying for idle seats. Seat plans such as ChatGPT Business charge for every purchased seat, assigned or not; check usage before renewing.
• Ignoring price changes. Introductory rates end and models retire. Our guide to managing LLM costs covers caching, batching and model routing.
• Testing on one example. Real enquiries vary in length; cost the 90th percentile, not the average.

Check your work

• I recalculated the token totals and the monthly model cost by hand.
• My comparison includes build and upkeep hours, not just model fees.
• I know whether my workflow needs to run inside our own systems.
• I checked current prices on the vendors’ own pages and dated them.
• I can explain in two sentences why I chose seats, API or both.

Where to go next

Next in the pathway: understanding AI tool permissions, which matters as soon as either route connects to your inbox or files. For choosing between plans, see choosing business AI subscriptions and our AI tools pricing for New Zealand businesses; to budget the whole programme, creating an AI budget. Run your own numbers in the AI cost calculator, and if you need help picking models for an API build, see AI model selection.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.