Guide · AI running costs
Building something is a one-off cost. Running it is three separate monthly bills. Here is what each one is, what it comes to in New Zealand dollars, and what makes it blow out.
Running AI in a business comes down to three separate monthly bills, and they behave differently. People mix them up, which is how you end up comparing a per-seat price with a per-token price and getting nowhere.
A business that only uses assistants pays the first bill. A business with an automation and no staff seats pays the second and third. Most end up with all three.
A business seat costs roughly NZ$30 to $45 per person per month. Prices were checked on 28 September 2026 and exclude GST. US-dollar plans are converted at about US$1 = NZ$1.75.
| Plan | Per person per month | Notes |
|---|---|---|
| ChatGPT Business | US$20 annual / US$25 monthly (about NZ$35 / $44) | Standard seat. |
| Claude Team | US$20 annual / US$25 monthly (about NZ$35 / $44) | Standard seat, includes Claude Code. |
| Microsoft 365 Copilot Business | NZ$34 annual / NZ$40.80 monthly | An add-on to a Microsoft 365 plan you already have. A promotional annual price of NZ$29.14 was running when checked. |
| Gemini in Google Workspace | Included from NZ$21 (Business Standard) | No extra charge if you are already on that plan or above. |
There is more on choosing between them in ChatGPT, Claude, Copilot or Gemini. The cost trap with seats is buying them for everyone on day one. Start with the people who asked. Pay month to month for the first two or three months, look at who is actually using it, then move those people to annual billing and cancel the rest.
A token is a piece of a word, and it is how AI models charge when a system uses them directly. As a rule of thumb, 1,000 tokens is about 750 words. You pay one rate for what the model reads (your instructions and the document) and a higher rate for what it writes back.
These are Anthropic's published prices on 6 October 2026, in US dollars per million tokens. OpenAI's and Google's sit in a similar range, with their smallest models cheaper again.
| Model size | Reading (input) | Writing (output) | Good for |
|---|---|---|---|
| Small and fast (Claude Haiku 4.5) | US$1 | US$5 | Sorting, extracting fields, simple replies. |
| Mid-range (Claude Sonnet 5.5) | US$2 | US$10 | Most business work: drafting, checking, summarising. |
| Large (Claude Opus 5.5) | US$4 | US$20 | Long documents and harder reasoning. |
A million tokens sounds abstract, so here is a real shape. Reading a typed one-page supplier invoice and returning the fields takes about 1,700 tokens in (the invoice plus the instructions) and about 300 out. On the small model that is roughly a third of a US cent. A thousand invoices a month comes to about US$3, or NZ$6. On the mid-range model it is about NZ$11. Scans and photos cost a bit more to read than typed PDFs.
That is why I describe usage as a few dollars to a few tens of dollars a month for a small business. I pass it on at cost.
A deep or thinking mode costs more because it uses a larger model and writes out its working before it answers. That working is billed as output, which is the expensive kind of token. It is also slower, sometimes by minutes.
A simple rule for staff: use the fast mode for drafting, summarising and looking things up, and the deep mode when the answer needs working out, like comparing three long contracts or finding why a spreadsheet does not reconcile. On a seat plan the deep mode is included, but it eats into the usage limit faster. In an automation you pay for every token of it, so I only switch it on for the steps that need it.
Anything custom-built needs somewhere to run and someone watching it, and that starts at NZ$99 a month. The cloud bill itself is small at this scale, often a few dollars. Most of what the plan pays for is attention: an alert when something fails, a fix the same day during New Zealand business hours, and updates when one of the systems it connects to changes.
The code and the cloud account are yours, so this is a service you can cancel, not a licence you are locked into.
A ten-person business with everyone on a seat and one automation running would spend roughly NZ$450 to $500 a month. This is an example to show how the three bills add up. It is not a quote.
| Bill | How it is worked out | Per month |
|---|---|---|
| Seats | 10 people on an annual business plan at about NZ$35 each | NZ$350 |
| Usage | One invoice workflow, about 1,000 documents | NZ$10 to $30 |
| Hosting and support | Monthly plan for the workflow | NZ$99 |
| Total | Excluding GST | About NZ$460 to $480 |
Seats are three-quarters of that. If only four of the ten use the assistant regularly, the same business spends about NZ$260. The one-off cost of building the workflow is separate and is covered on the pricing page.
Usage bills go wrong in four predictable ways, and each has a simple guard.
Everything I build ships with those caps switched on.
ChatGPT Business costs US$20 per person per month on annual billing or US$25 month to month, which is about NZ$35 or NZ$44 at US$1 = NZ$1.75, excluding GST. Claude Team is priced the same. Microsoft 365 Copilot and Gemini in Google Workspace are priced as add-ons or bundles to software you may already pay for.
A token is a piece of a word, and it is the unit AI models charge by. As a rule of thumb, 1,000 tokens is about 750 words. Models charge one rate for the text they read and a higher rate for the text they write.
Because it does more work. Deeper modes usually run a larger model and have it write out its reasoning before answering, and that working is charged as output, the more expensive kind of token. They are also slower. Use them for analysis and planning, not for everyday drafting.
Start with the people who asked for it, not the whole company. Pay month to month for the first two or three months, check the usage report, then move the regular users to annual billing and cancel the rest. Review the list every quarter.
The price of a given level of capability has kept falling. For example, Anthropic's current mid-range model costs US$2 per million tokens read and US$10 written, down from US$3 and US$15 for the previous generation. Seat prices have been steadier. Do not hold off a project waiting for cheaper models, because usage is normally the smallest of the three bills.
Tell me how many people you have and what you want to automate, and a free 30-minute call is enough to put real figures against all three bills. Discovery is always free.
Book a free call