What an AI agent really costs to run
The price of AI has fallen more than 280-fold, yet AI bills are rising. Agents use far more of it per task, and the model is only one line in the budget. Here is how to price an agent before you build it.
The short version
- The cost of querying a capable AI model fell from US$20 to US$0.07 per million tokens in about two years, a drop of more than 280-fold 1.
- Bills are rising anyway. Gartner says agents need 5 to 30 times more tokens per task than a chatbot, and expects overall inference costs to increase 2.
- Gartner also predicts more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs and unclear business value 3.
- The fix is to budget per finished task, across five cost lines, before the build starts. The model is often the smallest of the five.
Tokens got cheap. Agents did not.
AI models charge by the token, a unit of text a little shorter than a word. Stanford's AI Index tracked what it cost to query a model as capable as GPT-3.5: US$20 per million tokens in November 2022, and US$0.07 by October 2024 1. Few prices in business have ever fallen that fast.
So it is reasonable to ask why finance teams are suddenly paying attention. The answer is in how agents work. A chatbot answers one question with one call to the model. An agent reads a document, decides what to do, calls a system, checks the result and tries again, and at every step it sends the model the instructions and history so far. Gartner puts the difference at 5 to 30 times more tokens per task 2.
Gartner's forecast is blunt about where that leads. It expects the cost to providers of running large models to fall by more than 90% between 2025 and 2030, and it does not expect those savings to be fully passed on to customers. Because token use is rising faster than token prices are falling, it expects overall inference costs to go up 2.
The people who manage technology budgets have noticed. The FinOps Foundation's 2026 survey drew 1,192 respondents and reports that 98% now manage AI spend. Two years earlier it was 31% 4.
Price the finished task, not the token
A price per million tokens tells you almost nothing about what an agent will cost. The useful number is the cost of one finished task: one invoice processed, one enquiry answered, one report produced. That number can be estimated before a build and measured during a pilot.
Take an agent that processes supplier invoices. Assume each invoice takes six calls to the model, and each call sends about 10,000 tokens of instructions, document text and history. That is 60,000 tokens in and, say, 4,000 tokens out per invoice. These volumes are illustrative; your own will differ, which is the point of measuring them.
Priced on a mid-range model at published rates, the arithmetic looks like this 5.
At 1,000 invoices a month, that is US$160. Token volumes are illustrative. We use one provider's price list so the arithmetic can be checked; the method is the same for any provider.
Sixteen US cents an invoice is small next to the salary cost of the people who key, check and approve invoices today. But the same task on a different model gives a very different answer. On the same provider's smallest current model, the invoice costs under one cent. On its most expensive, it costs 64 cents 5.
That is an 80-fold spread for the same work. Choosing the right model for each step is the largest single lever on the model bill, and it is a design decision, not a procurement one. Most invoices do not need the most capable model. The unusual ones might.
Move the sliders to match a task in your business: how often it runs, how much the agent has to read each time, and how much it writes back. It starts on the invoice example above.
Is your AI providing value? No one can answer that question yet.
A practitioner quoted in the State of FinOps 2026 report
The model is one line of five
The reason projects are cancelled for cost is rarely the token bill on its own. It is that nobody wrote down the other lines. An agent that runs in production has five, and they behave differently as volume grows.
- Model usageScales with volume
Tokens in and out, per task. Estimated before the build and measured in the pilot 2, 5.
- Platform and integrationFixed each month
Hosting, the connections into your systems, and any licences those systems need for an agent to use them.
- People in the loopScales with exceptions
The time staff spend approving and correcting. Often the largest running cost at the start, and the one that should fall as the agent improves.
- Monitoring and upkeepFixed each month
Someone watching quality and cost, and updating the agent when a source system, a form or a model changes.
- Build and changeOne-off, then occasional
The first build, and each new process or system added later.
Four ways to bring the bill down
Right-size the model. Use a small model for reading and sorting, and a capable one only where judgement is needed. On current prices the gap between tiers is as much as 80-fold 5.
Stop paying to repeat yourself. Most of what an agent sends on each call is the same instructions as last time. Providers discount repeated input heavily when it is cached: on the price list used above, to a tenth of the normal input price or less 5.
Batch what is not urgent. Work that can wait, such as overnight reconciliations or month-end packs, can run through batch processing at half price 5.
Cap the loop. Set a limit on how many steps an agent may take on one task before it hands over to a person. It protects quality as well as cost, because an agent that is going round in circles is usually stuck.
Budgeting by the token
- A monthly AI allowance, set by guesswork
- One model for every step
- Cost reviewed when the invoice arrives
- No link between spend and work done
- Running costs discovered after go-live
Budgeting by the finished task
- A cost per invoice, enquiry or report, agreed up front
- The right model for each step
- Cost per task on a dashboard someone reads
- Spend set against the baseline cost of the task today
- All five lines estimated before the build starts
What this means in Australia and New Zealand
Australian organisations are spending more carefully than their peers. In Deloitte's 2026 survey of more than 3,000 senior leaders, 65% of Australian respondents planned to increase AI investment in the next financial year, against a global average of 84%. Yet 69% of Australian organisations were already using agentic AI, and only 22% reported a highly advanced model for governing agents 6.
That combination, agents in use, budgets tight and governance thin, is where unplanned costs come from. Two local points are worth adding to any budget on either side of the Tasman.
First, model prices such as those above are published in US dollars 5. An agent's model bill moves with the exchange rate, so a budget set in Australian or New Zealand dollars needs some room.
Second, ask where the model runs. If your risk or privacy position requires processing in Australia or New Zealand, confirm that the model you have priced is available there, and whether the price is the same.
A note on what you are being sold
Gartner estimates that of the thousands of vendors now selling agentic AI, only about 130 offer genuinely agentic products. It calls the rest agent washing: assistants and chatbots relabelled 3. That matters for cost, because a relabelled chatbot is typically priced per seat, while an agent should be priced by the work it completes.
The questions below work on any proposal, including ours.
Questions to ask before you approve the budget
If a proposal cannot answer these, the cost is not yet known.
- What does one finished task cost on this design, and how was that measured?
- Which model handles each step, and why that one?
- What does the task cost us today, in staff time and errors?
- How many tasks need a person to step in, and what is the plan to reduce that?
- What are the fixed monthly costs, and who pays them after handover?
- What limit stops a runaway agent, and who is told when it is hit?
Want these numbers for your own processes?
Our AI Opportunity Assessment costs the two or three automations most likely to pay back, across all five budget lines, against what each process costs you today.
See the AI Opportunity AssessmentSources
- Stanford HAI, AI Index 2025: State of AI in 10 Charts
- AIwire, Gartner forecasts 90% drop in LLM inference costs by 2030 (25 March 2026), reporting Gartner's press release of the same date
- Reuters via The Star, Over 40% of agentic AI projects will be scrapped by 2027, Gartner says (25 June 2025)
- FinOps Foundation, State of FinOps 2026
- Anthropic, Claude API pricing (read 10 November 2026)
- Deloitte Australia, Australian organisations lag global peers in realising AI's transformational potential (12 February 2026)