Most AI product pricing pages now use credits instead of clear units like dollars, minutes, or document counts. Credit value varies by vendor, with no industry standard. As a result, two products offering “500 credits a month” may deliver very different amounts of actual work, since each company defines its own credit system.
This is not a design oversight. The industry is billing for tokens, a unit of text processing that was unfamiliar outside research until recently. Tokens—whether input, output, cached, or reasoning—are often priced differently within the same product. When users see “1,240 tokens consumed” on a bill, the term refers to a vendor-created billing unit, not a measure of completed work.
Why the Meter Feels Rigged
The situation becomes more complex in practice. Within organizations or households, a few heavy users often consume most of the credits, a common pattern in metered services. Vendors call this “expansion revenue,” while buyers may face unexpected charges and must monitor usage through dashboards and approval processes. Notably, when researchers surveyed technical buyers, most preferred pricing based on completed work rather than tokens or credits.
“A token is a unit invented by the people billing you—for reasons that have nothing to do with the work you actually got done.”
So how do you budget for something this slippery? A few habits can help. Set a strict spend cap instead of trying to estimate token counts in advance, and nobody does that math reliably, including the vendors. Favor tools that price around finished work. To manage unpredictable costs, set a firm spending cap rather than estimating token usage, since even vendors struggle to predict it accurately. Choose tools that price based on completed work instead of raw consumption. Reserve advanced reasoning modes for tasks that require them, since a simple query can escalate in cost if it triggers unnecessary reasoning. Be cautious with “unlimited” plans; analyses show that a $20 monthly subscriber can cost providers $18 to $25 in compute during high usage, an unsustainable gap. Education researchers are already documenting a new kind of “AI divide,” distinct from the old dial-up-versus-broadband gap, one where even students with equal access split apart based on who has the literacy to use these tools well. That’s a real and measurable concern, not a hypothetical one.
However, AI costs are declining more rapidly than almost any previous technology. In late 2022, GPT-4-level performance cost about $20 per million tokens; by early 2026, similar performance cost less than one dollar. Researchers note this price drop surpasses those seen in computing and bandwidth during previous technology booms. As a result, AI that is unaffordable today may become inexpensive or free within two years. While opinions differ on the fairness of this transition, it is widely seen as temporary. Governments and cloud providers are already investing in subsidized compute to accelerate this shift.
Why AI Can’t Just Copy Google’s Free Playbook
A common question is why ChatGPT cannot adopt Google Search’s free, ad-supported model. The reason is economic, not strategic. Google Search relies on a pre-built index, with the main expense—crawling and indexing—spread across billions of queries, making each search inexpensive. In contrast, AI-generated answers are computed individually for each request.
Estimates place ChatGPT’s cost per query at approximately 0.36 cents, compared to Google’s 1.61 cents in ad revenue per search. Replacing all Google searches with AI-generated responses would eliminate tens of billions in operating income. Even Google’s AI search features cost about ten times more per query than standard search.
A second challenge is that traditional search results pages accommodate multiple paid links alongside organic results, driving advertising revenue. AI-generated answers consolidate information into a single response, reducing available advertising space.
So Who’s Going to Pay for This?
The industry is exploring multiple approaches, none of which resemble a traditional Google results page. OpenAI allows brands to integrate product information directly into ChatGPT’s responses. Some companies are testing affiliate models, earning a share of purchases made through AI recommendations. Perplexity lets brands sponsor suggested follow-up questions, while the AI generates the answer. Microsoft’s Copilot offers in-chat checkout and has reported higher click-through and conversion rates than traditional search ads for certain placements.
OpenAI’s early advertising pilot generated over $100 million in annualized revenue within six weeks, with projections reaching tens of billions by 2030. However, this remains small compared to Google’s $295 billion in ad revenue in 2025; even optimistic AI advertising figures are minor by comparison. Privacy concerns also remain: users often share sensitive information with chatbots, and most surveyed would trust AI answers less if ads were included. A former OpenAI researcher publicly resigned over this issue, cautioning that the company could repeat past mistakes seen on other platforms with sensitive data.
AI advertising, even under OpenAI’s own optimistic 2030 projection, is still dwarfed by what Google earns from search ads today.
The Quiet Force That Might Actually Fix This
There is reason for optimism: engineers are significantly reducing operational costs. Techniques such as distillation can reduce hardware requirements by up to eight times by compressing large models into smaller, efficient versions. Caching, or reusing answers to similar questions, can lower costs by up to 90% for repetitive tasks. Directing simple queries to less expensive models and reserving advanced models for complex problems further reduces average costs. Major AI providers have already lowered their prices to reflect these improvements.
However, AI products are becoming more agentic, often making many model calls in the background to complete a single task. While the cost per call is decreasing, the number of calls per task is increasing. The balance between these trends will likely determine future AI costs more than any pricing announcement. Currently, engineering efficiency appears to be outpacing increased usage.
