Running an AI agent costs far less than most quotes imply. An agent drafting 240 quote descriptions a month, sending 8,000 tokens in and 700 out per job, costs about USD 2.76 a month on Claude Haiku 4.5, roughly NZD 4.75. The build is the expense. Context length and retries are what move the running bill.
Tokens, in the only definition you need
A token is the chunk of text a model charges you for. In ordinary English it works out at about three quarters of a word, so a thousand words of job notes is roughly 1,300 tokens. Common words are one token each, longer or unusual ones split into two or three, and punctuation and spaces count too. You do not need to be more precise than that to budget.
Every provider publishes a price per million tokens, split two ways. Input is everything you send: your instructions, the reference material, the job notes, and the whole conversation so far. Output is everything the model writes back. Output costs several times more than input on every model I have priced, which is the opposite of what people assume, because the output is usually the short bit.
Those two numbers, multiplied by how often the thing runs, are the entire bill. There is no seat licence underneath it and no minimum. An agent that gets used twice in a month costs you the price of two calls.
Published rates, as at 8 September 2026
These come from the providers' own pricing pages, read on 8 September 2026. Rates change and models get renamed, so treat this table as a snapshot and check the page before you commit to a number: claude.com/pricing and OpenAI's API pricing.
| Model | Input, USD / million | Output, USD / million |
|---|---|---|
| GPT-5-nano | 0.05 | 0.40 |
| Claude Haiku 4.5 | 1.00 | 5.00 |
| Claude Sonnet 5 | 2.00 | 10.00 |
| GPT-5.6 Sol | 4.00 | 20.00 |
| Claude Opus 5 | 5.00 | 25.00 |
Those are US dollars. This site converts at NZD 1.72 to the US dollar, the rate a US dollar cost in late July 2026, the same convention used on the NZ trades software price index. The rate moves, and a converted figure is a guide rather than a quote.
Worked example: a quote-drafting agent for six people
Take a plumbing firm with six on the tools. The agent reads the site notes off the job card and drafts the description that goes on the quote. It runs 12 times a working day, 20 working days a month, so 240 calls.
Each call sends the standing instructions, a trimmed extract of the price book, and the job notes. Call that 8,000 tokens of input. It writes back a paragraph and a line-item list, about 700 tokens of output.
- Input: 240 calls x 8,000 tokens = 1,920,000 tokens, or 1.92 million.
- Output: 240 calls x 700 tokens = 168,000 tokens, or 0.168 million.
- On Claude Haiku 4.5: 1.92 x USD 1.00 = USD 1.92 in, plus 0.168 x USD 5.00 = USD 0.84 out.
- Monthly total: USD 2.76, or about NZD 4.75 at 1.72.
Swap the model and the arithmetic is identical, only the rate changes. Same 240 calls, same tokens:
| Model | Monthly USD | Monthly NZD at 1.72 |
|---|---|---|
| GPT-5-nano | 0.16 | 0.28 |
| Claude Haiku 4.5 | 2.76 | 4.75 |
| Claude Sonnet 5 | 5.52 | 9.49 |
| GPT-5.6 Sol | 11.04 | 18.99 |
| Claude Opus 5 | 13.80 | 23.74 |
Every one of those numbers is smaller than a single seat of the job management software the firm already pays for. Which is the point of doing the sum: the token bill is rarely the thing that kills an agent project. Build time, review time and the work of keeping the price book current are the real costs, and I put numbers on those in what custom automation costs. The same arithmetic applies when the agent answers the phone instead of drafting quotes; AI phone agents in NZ works through what that job costs and where it falls short.
Context length is the first thing that moves it
The easy mistake is to feed the model everything and let it work out what matters. It reads well in a demo and it multiplies your input line item by whatever factor you were careless by.
Say the agent stops using a trimmed extract and gets handed the full 40,000-token price book on every call. Input goes from 1.92 million tokens a month to 9.6 million. Swapping an 8,000-token context for a 40,000-token one takes that same 240-call month on Claude Haiku 4.5 from about USD 1.92 to about USD 9.60 in input charges, five times the bill for the same answers, at the rates published on claude.com/pricing as at 8 September 2026.
The fix is boring and effective. Retrieve the six line items the job needs instead of the whole book. Cap how much conversation history gets carried forward. Where the same block of text goes in on every call, check whether the provider offers caching on it, because cached input is billed at a fraction of the standard rate on both providers above.
Retries are the second, and they hide
A retry resends the whole context. When a tool call fails, a JSON response comes back malformed, or a validation step rejects the answer, the agent runs the call again from the top and you pay for the input twice.
One retry on one call in five adds about 20% to the month: USD 2.76 becomes USD 3.31 in the example above. That is survivable. What is not survivable is an agent stuck in a loop overnight because a supplier API started returning a 500, retrying every thirty seconds with a 40,000-token payload attached. Nothing in the token price warns you about that. A hard retry limit, a per-day spend cap, and an alert when the day's usage crosses a threshold you set are the guardrails, and they belong in the build rather than in the post-mortem.
The other quiet multiplier is an agent that gets things wrong and has to be asked again by a person. I wrote up how that happened to me in why my agent kept giving confident wrong answers. Every re-ask is another full-price call, and the human time attached to it costs more than the tokens ever will.
A real pipeline, with a real number
The Canterbury lead generator is my own agent pipeline: it finds Canterbury businesses, crawls their sites, and scores them. Its page says the whole build cost about $8 in API credits, running on Claude Haiku, and that one overnight run after the July rebuild found 1,579 leads, verified 649 emails and scored 355 of them hot.
That is a pipeline reading thousands of web pages, which is about as input-heavy as a small-business agent gets, and the model bill is still lunch money. The expensive parts were the 290 unit tests and the weeks of tuning that stopped it scoring rubbish as hot.
Microsoft bills the same work differently
If the agent lives in Copilot Studio rather than on a raw API, none of the token arithmetic above applies to your invoice. Microsoft meters by Copilot Credit, and the credit count depends on what the agent did to answer rather than how much text moved. A scripted reply draws about 1 credit, a generative one about 2, and grounding the answer in your own SharePoint or Outlook data adds roughly 10 on top.
That flips the optimisation. On a token meter you shorten the context. On a credit meter you reduce how often the agent grounds, because a two-sentence answer that reads your tenant costs ten times a two-sentence answer that does not. The full breakdown, including how to cap an agent before the invoice arrives, is in the Copilot Studio message meter.
When you do not need this arithmetic
Skip the whole exercise if the agent runs a handful of times a day. Below about 50 calls a day on a small model, the monthly bill lands under NZD 20 and modelling it costs more of your time than it saves. Turn it on, watch the first invoice, and move on.
Skip it too if what you want is a chat assistant for your staff rather than an agent doing a job unattended. That is a per-seat purchase with a fixed monthly price and no token meter at all, and the comparison you want there is Copilot Studio against ChatGPT Business.
Do the arithmetic when the agent will run unattended, when it reads large documents, or when someone has quoted you a monthly running cost with three digits in it. In that last case the number usually covers hosting, monitoring and support, which is fair, and the tokens inside it are a rounding error worth seeing separately.
Before you sign off on an agent
If someone has put a running cost in front of you and you want it checked against the published rates, book a free 30-minute call and I will do the sum with you on your own call volumes. The ROI calculator covers the other half, what the work is worth once the agent takes it off someone's desk.