Companies are finding it almost impossible to budget for AI spending, the Wall Street Journal reported on October 5. A recent study found that only 11 per cent of nearly 400 surveyed businesses could accurately forecast their AI costs.

AI differs from standard software because it behaves more like a human worker. It takes actions, makes decisions and makes mistakes, all while the billing clock runs on tokens.

More advanced models usually cost more per token but can work more efficiently. Researchers from Stanford, Carnegie Mellon, UC Berkeley and Microsoft Research tested models on more than 6,800 tasks.

In 32 per cent of cases, lower-priced models actually cost more than higher-priced ones. In one example, Google’s Gemini 3 Flash ran up $14 of token use before failing, while the pricier 3.1 Pro finished the same task for $1.

The researchers’ advice is clear: price alone should not be used to infer which model is actually cheaper for a given task.