The Token Tangle: Why AI Pricing Has Never Been Straightforward

When you tap a free AI assistant on a phone screen you might think a few words in your pocket will serve as a fee. Real‑world costs for the research and infra that powers models such as ChatGPT, Claude and Gemini run in the billions, so the value you receive feels disproportionately large. These giants therefore launch paid tiers offering extra functions for coding, billing and advanced analytics, in the hope that deeper use will bring back the capital invested.

But setting a price for AI products is far more difficult than a simple subscription. Simon Gooch from Saviynt says that even a 12‑month contract is hard to justify when the number of tokens a model consumes changes with each prompt. Tokens are the building blocks of LLMs and the amount of text that goes in or comes out is measured in these units. A small change in wording can lead to a very different output, changing the token count and thus the bill.

Goldman Sachs predicts token consumption will multiply 24 times from 2026 to 2030, reaching about 120 quadrillion tokens a month, as companies shift toward fully autonomous AI agents. Yet the price per token has dropped, while the volume demanded has risen sharply, creating a mismatch that keeps businesses chasing the curb of their spending. As companies build AI inside products for thousands of users, the hidden cost can balloon, especially when AI is used for testing, security or adding guard rails.

Will Venters, a professor at the London School of Economics, warns that staff projects can burn tokens faster than expected, and that “cost is non–deterministic”. Some firms have turned to flat‑fee personal accounts to hide usage, a practice the major vendors will likely oppose once shareholders demand clear profit signals. Matches of toner by smaller organizations may shrink before larger vendors clamp down on those accounts.

Companies are also learning to choose the right kind of AI model and to design precise prompts. Rosetta fraud teller sees that harvesting tokens can cost less when the model knows exactly what output is needed. CFOs of accounting firms caution that “you must not send an AI participant to collect groceries without explicit guidelines, because the result can be expensive and unpredictable.”

Despite these challenges, many businesses believe they benefit from token use. Venters argues that the calculus is not identical to a calculator: the more tokens a model receives, the better the result, but the higher the price. The trade‑off pushes leaders to decide how to pass on AI costs to customers, considering options like bundling incidents or charging by results.

In the near term, volatility will persist. Bill Peterson from Sumo Logic notes that AI pricing is still under debate internally and will likely shift if underlying LLM vendors revise their own pricing structures. “Variable pricing, changing every few months, leaves customers uncertain,” he adds. As the AI economy matures, companies will need new budgeting tools and clearer transparency on token consumption to keep the ecosystem stable.

The future of AI pricing is still being tested by businesses, especially as the rapid rise of token usage creates a hidden cost that neither model alone nor data provider alone controls. The industry will need a balance of product innovation and transparent billing to sustain the benefits of AI without derailing fiscal stability.