AI pricing puzzle: why firms struggle to cost tokens, agents and subscriptions

New Delhi: Free versions of ChatGPT, Claude and Gemini have made powerful AI feel like a bargain for consumers and small businesses. Behind these interfaces, however, lie large language models (LLMs) built with hundreds of billions of dollars in investment by companies such as Microsoft, Google and Anthropic. To recoup costs, these firms sell premium tiers with advanced features, while third‑party vendors build specialised AI agent services on top of the same models. Setting stable prices for these services has proved unexpectedly difficult.

Tokens: the invisible currency

When a user asks an LLM to draft text, generate code or automate a task, the request is converted into tokens—mathematical units the model processes. The response is also produced in tokens and then turned back into readable output. Costs are tied to token usage, but consumption is hard to predict. Small changes in prompts can lead to very different outputs, and the same prompt may yield different results across runs or models. In agentic systems, where multiple AI agents collaborate to make decisions and take actions, token use multiplies and becomes even less predictable.

Soaring usage, uncertain bills

Token prices have fallen sharply in recent years, yet total consumption is surging. Goldman Sachs forecasts that monthly token use could rise 24‑fold between 2026 and 2030, reaching 120 quadrillion tokens as companies shift to AI agents. Many organisations, however, have only a vague idea of how many tokens they are burning through until they hit limits or receive large monthly bills. Reports suggest Microsoft has curbed engineers’ use of some third‑party coding tools, while Uber allegedly exhausted its annual AI coding token budget within months.

Workarounds and coming crackdowns

Some smaller firms currently rely on flat‑fee personal accounts to keep costs down, a practice large vendors tolerate for now but are unlikely to sustain under shareholder pressure for profits. Experts warn that as platforms seek profitability, they will tighten access and revise pricing. Companies are being urged to choose models more carefully and craft precise prompts to avoid wasteful token use.

Passing costs to customers

For software vendors embedding AI into products, costs can balloon quickly as tokens are needed for development, testing, security and guardrails. Pricing options under discussion include across‑the‑board price hikes, pay‑for‑results models, or bundled incident packages. Yet any structure risks being disrupted if major LLM providers change their own pricing every few months, leaving customers without a stable basis for budgeting.