AI may be getting cheaper to run, but that does not mean companies are spending less on it.

The reason is simple: businesses are using AI far more often, and newer systems can perform many more steps on their own. A chatbot might answer one question and stop. An AI agent can read documents, write code, call software tools, check its work and try again. Each step can consume more tokens.

Tokens are the basic units used to process text and other information inside large language models. Providers charge customers according to how many tokens their systems process, although pricing varies between models and platforms. API pricing shows how input, output and cached tokens can carry different costs.

That creates a difficult problem for companies building AI products. They need to know how much an AI service will cost before they can decide what to charge customers. The trouble is that token use can change dramatically depending on the task, the model and the number of steps an AI agent takes.

Why AI Token Costs Are Hard to Predict

Traditional software is usually easier to budget. A company buys a license, pays for cloud capacity or agrees to a subscription. Costs may increase as usage grows, but the basic structure is relatively easy to understand.

AI can behave differently.

A simple request might require only a small amount of processing. A complicated task can trigger a much longer exchange between the user and the model. An AI agent may also make several tool calls, review its own work and generate additional instructions before completing the job.

That makes token consumption less predictable.

The cost of individual tokens has fallen sharply as AI models and computing infrastructure have improved. At the same time, businesses are using AI in more places. Coding, customer service, security, research and internal automation can all generate large volumes of model activity.

Goldman Sachs has projected that monthly token consumption could rise more than 24 times by 2030 as agentic AI becomes more widely used. Sachs AI agents analysis

The result is an unusual economic equation: each token may become cheaper while the total number of tokens keeps rising.

For businesses, that distinction matters.

A company might reduce its cost per million tokens and still see its overall AI bill increase because employees and automated agents are consuming far more of them.

AI Agents Could Make the Pricing Problem Bigger

The challenge becomes harder when companies move from chatbots to AI agents.

An ordinary chatbot generally waits for a prompt and produces an answer. An agent can take action. It may search a database, write a piece of code, test that code, identify a problem and run another process.

Every additional step can increase consumption.

Research into agentic coding tasks has found that token use can vary widely between similar tasks. The same study also found that higher token consumption does not automatically produce better results, suggesting that simply giving an AI system more room to reason is not always an efficient strategy.

That is why companies are beginning to pay closer attention to model selection and workflow design. A powerful model may be unnecessary for a routine task. A smaller or cheaper model could handle it adequately while keeping costs under control.

Model pricing is also becoming more complicated. Providers can charge different rates for input and output tokens, while cached requests and batch processing may have separate prices. Anthropic, for example, publishes different rates depending on the model and processing method. model pricing

For companies selling AI-powered software, this creates another problem: customers generally want predictable bills.

A software buyer may accept a usage-based model for certain services. But if the price changes every month because an AI agent consumed an unpredictable number of tokens, budgeting becomes much harder.

Who Pays When AI Usage Keeps Growing?

Companies developing AI products have several options.

They can charge customers according to usage. They can create monthly plans with limits. They can bundle a fixed number of AI operations into a subscription. Or they can charge according to the result produced by the system rather than the amount of computation used behind the scenes.

None of those models is perfect.

Usage-based pricing protects the seller when consumption rises, but customers may find the bills difficult to predict. Flat subscriptions are easier to sell, but they can expose the provider to heavy users who consume far more computing resources than expected.

That tension is likely to become more important as AI agents move deeper into business operations.

Companies may also need better tools to monitor where their tokens are going. A large AI bill may not come from one obvious application. It can build up through testing, security checks, automated reviews, repeated prompts and background agents running across different departments.

The economics are changing quickly enough that today’s pricing model may not survive for long. Falling token prices can encourage more usage, while greater usage can push total spending higher. At the same time, improvements in model efficiency could change which systems businesses consider affordable.

For now, the basic problem remains unresolved: AI providers can calculate the price of a token, but businesses still have a harder time calculating the value of the thousands or millions of tokens required to complete a real-world task.

That gap between cost per token and value per task is becoming one of the central problems in AI economics. on AI token economics.