AI looks cheap when you use it once. A single answer can cost a fraction of a cent. The bill arrives when that answer repeats across thousands of customers, teams and automated workflows.
To understand the cost of AI tokens, businesses need to know three things. They need to know what they are buying, how usage grows and which outcomes the spend produces.
What an AI Token Is and How It Is Priced
A token is a small unit of text, code, image or audio that an AI model reads or writes. In text, a token is often part of a word. This has nothing to do with cryptocurrency.
Tokens are fast becoming the unit businesses use to buy and consume AI. Deloitte’s CFO guide to AI token economics argues that tokens, not licences or headcount, are now the true unit of AI cost.
AI token pricing has three main parts:
- Input tokens cover everything the model reads. This includes the prompt, system instructions, retrieved documents and conversation history.
- Output tokens cover the response. They usually cost several times more than input.
- Cached tokens cover repeated content the model has already processed – providers such as Anthropic charge a lower rate for cached input.
A low price per token can hide the real cost. What matters is the total number of tokens used across every interaction, at scale.
Why Agentic AI Multiplies Consumption
A chatbot answers a prompt once. An agent works through a task in many steps.
Each step consumes tokens. Agents reason through a problem and retrieve documents. They call databases, APIs and business applications. Then they call the model again to check the result. Every call re-reads the instructions and history that came before.
Hidden drivers add more. Long prompts, growing conversation history, reranking, summarisation, evaluation and retries after errors all add tokens. Few of these show up in a demo.
Token prices are falling, but usage is growing faster. Oracle notes that AI costs can rise sharply when pilots move into production. Industry research shows some workflows now exceed $100,000 a year.
This is the pilot-to-production trap. A test with 50 users looks affordable. The same design can cost far more once it runs across customer service, internal workflows and workforce-wide token allowances.
Measure Cost per Outcome, Not Cost per Token
Cost per token tells you what you pay. It does not tell you what you get.
Measure AI unit economics by cost per resolved query, processed document, completed task or sale. Then ask a simple question. Does the value of the outcome justify the spend?
Count the related costs too. Orchestration, retrieval, storage, monitoring, security, human review, and rework all belong in the total.
A Simple Worked Example
A business handles 100,000 customer queries a month. We use illustrative prices. The premium model costs $3 per million input tokens and $15 per million output tokens. The smaller model costs $1 and $5. A human handles each unresolved query at $5.
| Basic chatbot (premium model) | Agent (premium model) | Agent (smaller model) | |
| Tokens per interaction | 1,500 in / 300 out | 60,000 in / 4,000 out | 48,000 in / 2,000 out |
| Cost per interaction | $0.009 | $0.24 | $0.058 |
| AI cost per month | $900 | $24,000 | $5,800 |
| Resolution rate | 40% | 85% | 83% |
| AI cost per resolution | $0.023 | $0.282 | $0.070 |
| Human escalation cost per month | $300,000 | $75,000 | $85,000 |
| Total monthly cost | $300,900 | $99,000 | $90,800 |
The chatbot has the lowest token bill and the highest total cost. The premium agent resolves the most queries. The smaller-model agent gives the best unit economics.
The largest model is not always the most commercially effective. A smaller model that resolves the query at lower cost often wins.
