Artificial intelligence subscriptions no longer behave like simple media services. Paying $20, $100 or $200 per month grants access to models, but consumption is not measured merely by time or number of uses. Modern AI offerings resemble a mix of mobile data, cloud storage and metered electricity: models, tokens, context windows, credits and reset times determine how much computationally produced "thinking" you receive for your money.
A token is the unit of text the model processes; the context window is the amount of text the model can consider at once. These resources are often limited and billed differently when usage involves longer sessions, larger inputs, or multi-step workflows that call external tools.
Metered usage and agents: when does extra payment kick in?
OpenAI, for example, describes agent-based usage for ChatGPT Plus and Pro subscribers: processes that work across multiple steps—invoking tools, modifying files, running code, or orchestrating subtasks—are counted against a shared agent usage quota, and when that quota is exhausted users must consume additional credits. The Codex documentation similarly notes that larger codebases, longer runs and extended sessions consume more quota.
Anthropic follows the same logic: Claude Pro and Claude Max subscribers can continue after hitting package limits by using consumption-based credits. These examples illustrate a broader trend: while simple, short interactions may become cheaper, deep, long-duration or agent-oriented workflows increasingly require separate metering and charges.
Cognitive capital: access alone does not democratize thinking
Access to AI does not automatically equal democratized thinking. Interfaces and tools may look uniform, but what they mean in practice depends heavily on a user's operational skillset—what I call "cognitive capital."
Cognitive capital is related to Pierre Bourdieu's concept of cultural capital, but it is narrower and more operational: it denotes the ability to form hypotheses, ask productive questions, seek evidence, detect overclaiming, verify sources, deconstruct an argument and rebuild a line of thought. These skills let a user turn AI into an accelerator, interlocutor, tutor, editor or fact-checker rather than merely a generator of answers.
People who lack this set of skills can become passive consumers: they may accept confidently phrased but incorrect outputs, fall prey to confirmation bias, or develop an illusion of competence. Therefore, beyond technical access, social trajectories—schooling, workplace practice, professional communities and reading and problem-solving habits—shape who can use AI critically and effectively.
Inequalities and a central paradox
AI creates a paradox: it simultaneously expands access and generates new inequalities. On one hand, people who previously could not afford editorial, tutoring or programming help may now use powerful language tools. On the other hand, users who cannot assess the quality of model outputs may end up worse off, because the systems communicate mistakes with great confidence.
Analysing AI consumption therefore requires more than comparing packages and prices. It also involves examining the everyday practices, institutional frameworks and educational opportunities that determine whether access translates into genuine cognitive capacity.
Practical consequences
Providers such as OpenAI and Anthropic technically measure and bill advanced usage; users must make operator-type decisions: which model to select, how to allocate their quota, what tasks to entrust to AI, when to perform manual checks, and when buying extra credits is worthwhile. These are not merely cost questions but questions of skill and judgment.
In sum, the pricing and metering practices of AI subscriptions are creating a quota-based cognitive infrastructure: a resource shaped not only by technology but also by social and pedagogical conditions.
