Tokens are the basic units artificial intelligence models use to process and generate information. Rather than reading complete words, AI systems divide text into small pieces, typically about four characters long. Prompts, generated responses, computer code, and even uploaded images are converted into tokens. Every token requires computing resources and therefore carries a cost, tells The New York Times (full article available to subscribers).
As AI models have become more sophisticated, token consumption has increased dramatically. Newer systems require advanced chips, greater computing infrastructure, and more energy. Yet consumers often do not see the full expense because companies offer free services or subscriptions priced below the actual cost of providing them while competing for users.
The difference between models can be substantial. In a New York Times test asking three generations of ChatGPT for the same road trip advice, the most advanced model consumed more than 14 times as many tokens as an earlier version. Token prices also vary because advanced models require newer hardware and considerably more energy.
Reasoning models add another layer of expense. These systems generate internal reasoning tokens before producing an answer, and most of that computational work remains invisible to users. This approach has improved AI performance on demanding problems such as mathematics and complex software development, but using advanced reasoning for routine tasks can waste resources.
The economics become increasingly important as AI adoption expands. Consumers accustomed to flat monthly subscriptions may eventually face higher prices if companies reduce subsidies and begin passing more infrastructure costs to users.
Businesses are also starting to scrutinize AI spending after initially encouraging employees to use the technology widely. Services such as OpenRouter have emerged to help organizations match models to appropriate tasks rather than automatically selecting the most powerful option.
Ultimately, every AI interaction has a cost. Whether users pay directly or not, growing token consumption carries financial, computing, infrastructure, and environmental consequences.
