Token management: what AI really costs, and how to tell if it pays.
AI is billed by the token. As businesses move from chat to agents, managing that bill, and proving what it buys, is becoming a real job.
The short answer
A token is the unit AI is billed in, roughly three-quarters of a word.
Prices per token keep falling, but AI agents use many more of them, so total bills are rising.
Token management means tracking that spend by task and comparing the cost of each completed outcome with what the same work costs you today.
Key facts.
- One token is about 4 characters, or three-quarters of a word, in English. 100 tokens is roughly 75 words.OpenAI, November 2026 · source
- Output tokens cost 4 to 6 times as much as input tokens at Anthropic, OpenAI and Google, usually 5 times. The most capable models cost about 100 times as much as the cheapest.Vendor pricing pages, checked 7 November 2026 · sources
- The price of reaching a fixed level of AI performance fell between 9 and 900 times a year, depending on the task.Epoch AI, March 2025 · source
- Google's models now process nearly 32 billion tokens a minute through its APIs, more than four times the figure a year earlier.Alphabet Q3 2026 earnings, 28 October 2026 · source
- AI agents typically use about 4 times the tokens of a chat, and multi-agent systems about 15 times.Anthropic, June 2025 · source
- Gartner predicts the inference cost of each agentic workflow will rise more than fivefold through 2028, even as token prices fall.Gartner, 17 August 2026 · source
- 98 percent of organisations in the State of FinOps 2026 survey now manage AI spend, up from 31 percent two years earlier.FinOps Foundation · source
What is a token?
A token is a small chunk of text, usually part of a word, and it is the unit every AI model reads, writes and bills in.
When you send a prompt, a large language model breaks it into tokens, a step called tokenisation. The words you type, the documents you attach, the instructions running in the background and the tools the AI can use all count as input tokens. Everything the model writes back counts as output tokens. Our guide to what AI is and how the types fit together covers the basics.
There is a third kind most people never see. Reasoning models "think" before they answer, and those thinking tokens are billed as output, even though you see only a summary or nothing at all. Anthropic and OpenAI both say so, and OpenAI notes a single response can use anywhere from a few hundred to tens of thousands of reasoning tokens.
Token counts also vary by model. Anthropic's pricing notes say its newer models produce about 30 percent more tokens for the same text. A price per token only means something next to the model it applies to, and how much text fits in one request is set by the model's context window.
What do tokens cost?
For a single task, very little. Each draft, summary or reply usually costs cents.
Published prices per million tokens, in US dollars, checked 7 November 2026:
| Model | Input | Output | Where it fits |
|---|---|---|---|
| Claude Fable 5.1 | $10 | $50 | Hardest, longest-running work |
| GPT-6 Astra | $10 | $50 | OpenAI's flagship |
| Claude Opus 5.5 | $4 | $20 | Complex reasoning |
| Claude Sonnet 5.5, GPT-6.1 Sol | $2 | $10 | Most everyday business work |
| Gemini 4 Argon | $2 | $10 | Google's flagship; doubles to $4 and $20 on 1 December 2026 |
| Claude Haiku 5.5, GPT-6 Luna | $0.10 | $0.50 | Simple, high-volume tasks |
Sources: Anthropic, OpenAI and Google pricing pages. Haiku 5.5 prompts over 100,000 tokens cost more.
Two patterns matter more than any single price. Output costs about five times input, so long answers and heavy "thinking" drive the bill. And the most capable models cost around 100 times the cheapest, so choosing the right model for each job is the biggest lever you control.
Why is token management becoming important?
Token prices are falling, but the number of tokens businesses use is rising faster: cheaper tokens, bigger bill.
Three shifts are behind it.
- Agents use far more tokens than chat. A chat is one question and one answer. An AI agent plans, calls tools, reads the results and tries again. Anthropic measured agents at about four times the tokens of a chat, and multi-agent systems at about fifteen. Connecting AI to your systems through MCP adds tool descriptions to every request too.
- Billing is moving from seats to usage. The flat monthly licence is giving way to pay as you go. Claude Enterprise charges a seat price plus usage at API rates. Microsoft Copilot now bills its more advanced experiences, such as Copilot Cowork, in Copilot Credits. ChatGPT Business and Enterprise use credits for agent work.
- Volume is climbing fast. Google reported nearly 32 billion tokens a minute in October 2026, up 40 percent in a quarter. At Microsoft's July 2026 earnings, its chief financial officer said customers will need to see and manage their token spend.
The result is a cost that grows quietly with use. Gartner's August 2026 forecast puts it plainly: the cost per agentic workflow will rise more than fivefold through 2028, even as the price of each token falls. It also expects over 40 percent of agentic AI projects to be cancelled by the end of 2027, with escalating costs among the reasons.
How do you attribute business value to token cost?
Measure the cost of each successful outcome, then compare it with what the same outcome costs you today.
Token counts on their own tell you nothing about value. The FinOps Foundation puts it simply: measure the use case, not the token. Four steps make that practical.
- Pick a unit of work. A drafted reply, a reconciled invoice, a completed report: something a person recognises as done.
- Work out the cost per successful outcome. Total AI spend for that workflow, divided by the outcomes that met your quality bar. Failed runs and re-runs count. The FinOps Foundation's example: $1,200 across 8,000 successful outcomes is $0.15 each.
- Add the human time still needed. Someone checks and sends the draft. That review time is part of the cost, and usually the larger part.
- Compare with today's cost. Time the task as it runs now, and value it at what that time costs you, not what you charge for it. The value is the saving, tied to a number you already track, such as response time, return on investment or cost to serve.
A worked example · Illustrative
A team answers 400 customer emails a week, and each takes 12 minutes today. With AI, a mid-tier model reads the email and the relevant policy, about 6,000 input tokens, and writes a reply, about 2,000 output tokens with its thinking included. A person checks and sends it in 3 minutes.
- Input: 6,000 tokens at $2 per million
- US$0.012
- Output: 2,000 tokens at $10 per million
- US$0.020
- Token cost per draft
- about US$0.03
- Token cost for 400 drafts a week
- about US$13
- Staff time back each week (9 minutes x 400)
- 60 hours
Illustrative figures, built from the published prices above. Your own token counts and timings will differ: measure them.
That is why the token bill is rarely the problem for one well-designed task. It becomes a problem when an agent loops, a flagship model does work a small one could, the same 50 page policy is re-sent with every request, or nobody can say which workflow the spend belongs to.
How do you manage token costs?
Five levers, in rough order of impact.
- Match the model to the job. Anthropic's own cost advice is a small model for simple tasks, the mid-tier for most production work, and the top model only for the hardest reasoning. Our comparison of AI tools sets out which plans include what.
- Cache what repeats. Policies, templates and instructions sent with every request can be cached. Anthropic charges a tenth of the normal input price for a cache read, or less on its newest models, and OpenAI and Google offer similar discounts.
- Batch what can wait. Overnight reports and bulk processing sent through the batch APIs cost 50 percent less at all three vendors.
- Cap the output and the thinking. Set limits on response length and reasoning effort where depth is not needed.
- Stop rework. The most expensive tokens are the ones spent building the wrong thing. A clear brief and a reviewed plan save more than any technical setting. Our free setup checklists for Claude and Copilot help the AI get it right first time.
Above all, give every workflow an owner and a budget, and watch spend by workflow, not by person. Admin consoles at Anthropic, OpenAI and Microsoft now support spend limits and usage reporting: see our guides to Claude, ChatGPT and Microsoft Copilot.
What is different for New Zealand businesses?
Three things: the currency, GST and where your data runs.
- Prices are in US dollars. Anthropic, OpenAI and Google price their APIs in US dollars, so your bill moves with the exchange rate. Microsoft prices Copilot Credits in New Zealand dollars, currently NZ$323.60 for 25,000 credits a month.
- GST on offshore services. Offshore suppliers of remote services generally charge New Zealand customers GST unless the customer tells them it is GST-registered, by giving its GST number or NZBN. See Inland Revenue's guidance on remote services.
- Data residency costs extra. Keeping processing in one region carries a surcharge: Anthropic charges 1.1 times for US-only inference, and OpenAI adds 10 percent for data residency. If your contracts require data sovereignty, budget for it, and check how the Privacy Act applies.
Not tax advice. This page explains the general position. For your organisation's GST position, talk to your accountant.
Our view.
Budget by workflow, not by seat. The per-seat era is ending. As AI moves from answering questions to doing whole jobs, the bill follows the work, and work you cannot see is spend you cannot manage.
The fix is the same discipline we bring to everything else: move a metric, not a use case. Every AI workflow should have an owner, a baseline, a number it is meant to move, and a cost per outcome someone looks at each month. Then the conversation changes from "the AI bill went up" to "we pay 15 cents to save 9 minutes, 2,000 times a month".
And treat agent cost the way a factory treats its line. Sampling, monitoring and a budget for assurance are part of running agents in production, not an optional extra. Our AI Operating System builds that in from the start.
What are the risks of unmanaged token spend?
Four come up again and again, and each has a simple guardrail.
- Runaway agents. An agent stuck in a loop can burn through a month's budget in hours. Set spend limits and alerts in the admin console, and keep a person approving what matters.
- Invisible spend. Usage on personal cards and unapproved tools never reaches the budget. See our guide to shadow AI.
- Price changes. Vendors change prices with little notice: Google's flagship model doubles in price on 1 December 2026. Review costs monthly.
- Measuring the wrong thing. Cost per token rewards cheap models that fail. Cost per successful outcome does not. Cheap answers that are wrong cost more, as our guide on why AI models hallucinate explains.
Where should you start?
Know where you stand, then measure the work before you measure the tokens.
- Find out where you stand. The free NZ AI Benchmark is 40 questions, takes about 20 minutes, and shows where your organisation stands and the next move.
- List your AI workflows and give each an owner. Include the ones staff have set up themselves.
- Time the task today. Without a baseline, there is no value to attribute.
- Turn on usage reporting and spend limits in each tool's admin console, and tag spend by workflow.
- Review cost per outcome monthly. Change the model, cache or batch where the numbers say so. For the governance behind it, the free AI Champions Committee Resource Pack has the register and templates, and our AI policy guide covers who may approve new tools.
Frequently asked questions.
How many words is a token?
In English, one token is roughly three-quarters of a word, or about four characters, so 1,000 tokens is around 750 words. Other languages, code and numbers can use more tokens per word, and different models count slightly differently.
Why is my AI bill going up if token prices are falling?
Because you are using far more tokens. Agents plan, call tools and retry, using several times the tokens of a chat. Reasoning models bill their hidden thinking as output. And as more people use AI for more work, volume grows faster than prices fall.
Do I pay for tokens on ChatGPT, Claude or Copilot business plans?
Increasingly, yes, for some features. Standard chat is usually covered by the seat price. But Claude Enterprise adds usage at API rates, Microsoft bills advanced Copilot features in Copilot Credits, and ChatGPT uses credits for agent work. Check each plan's current terms.
What is a good cost per task for AI?
There is no universal number. A good cost per task is one that is clearly lower than what the same task costs you today, including the human time to check the output. Measure cost per successful outcome, not cost per token, and compare it with your own baseline.
What is prompt caching?
Caching stores the parts of a request that repeat, such as a policy or a set of instructions, so the model does not reprocess them at full price each time. Anthropic, OpenAI and Google all discount cached input heavily, often by around 90 percent.
Do NZ businesses pay GST on AI services?
Offshore suppliers generally charge GST to New Zealand customers who are not GST-registered. If your business is registered, give the supplier your GST number or NZBN. This is general information, not tax advice: check your position with your accountant.
Sources.
- What are tokens and how to count them, OpenAI (help article; accessed 7 November 2026)
- Pricing, Anthropic (checked 7 November 2026)
- Thinking, steering and cost, Anthropic (documentation; accessed 7 November 2026)
- API pricing, OpenAI (checked 7 November 2026)
- Reasoning models, OpenAI (documentation; accessed 7 November 2026)
- Gemini API pricing, Google (checked 7 November 2026)
- LLM inference price trends, Epoch AI (March 2025; accessed 7 November 2026)
- Q3 2026 earnings remarks, Alphabet (28 October 2026; accessed 7 November 2026)
- FY26 Q4 earnings, Microsoft (29 July 2026; accessed 7 November 2026)
- How we built our multi-agent research system, Anthropic (June 2025; accessed 7 November 2026)
- AI inference costs per agentic workflow will increase more than fivefold through 2028, Gartner (17 August 2026; accessed 7 November 2026)
- Over 40 percent of agentic AI projects will be cancelled by end of 2027, Gartner (25 June 2025; accessed 7 November 2026)
- FinOps for AI overview, FinOps Foundation (accessed 7 November 2026)
- Claude plans and pricing, Anthropic (checked 7 November 2026)
- Microsoft Copilot pricing for business and Copilot Studio, Microsoft (checked 7 November 2026)
- ChatGPT rate card, credit-based pricing, OpenAI (help article; accessed 7 November 2026)
- Supplying remote services into New Zealand, Inland Revenue (accessed 7 November 2026)
Related topics
Keep going.
Your next step
Find out what your AI is really costing, and earning.
The free NZ AI Benchmark takes about 20 minutes and shows where your organisation stands. Or book a 45 minute Discovery Call with Justin or Ben to work out the cost and value of your first AI workflows.
Changelog: 7 November 2026, first published.