Big AI firms such as Microsoft, Google and Anthropic offer paid tiers of tools like ChatGPT, Claude and Gemini to recoup the huge sums spent developing them, but pricing these services is proving difficult. This matters because businesses building on top of these models, and third-party firms selling AI agents, cannot reliably predict or control their costs, making budgeting and long-term contracts problematic.

The core issue is "tokens", the units into which AI prompts and responses are broken down, whose consumption is unpredictable because outputs vary even for similar prompts, and multi-agent systems compound this further. Although the cost per token has fallen sharply, overall usage is soaring, with Goldman Sachs forecasting a 24-fold rise in token consumption between 2026 and 2030, to 120 quadrillion tokens a month. Firms including Microsoft and Uber have reportedly had to rein in or been caught out by runaway AI token spending, while some smaller organisations reportedly use flat-fee personal accounts to dodge costs, a workaround experts say cannot last once AI providers come under pressure to turn a profit.