An agentic AI system uses between 5 and 30 times more tokens to complete one task than a chatbot uses to answer one question, by Gartner's count. Neither the model nor the published per-token rate has to change for that to happen. Finishing a task means reading files, planning steps, calling other software, checking the result and often doing part of it over.
The tokens themselves keep getting cheaper to produce. Gartner forecast in March that by 2030 the cost to providers of running a model with a trillion parameters will be more than 90% below what it was in 2025. In the same forecast Gartner says those savings will not be passed on to business customers in full, and that token use is growing faster than token prices are falling, so what companies pay to run these systems goes up.
The per-token rate is the part a supplier publishes. The token count belongs to the buyer: how many steps a workflow takes, how much of a contract or a case file the model re-reads at each step, how often it tries again. A department can add no new users and sign no new contract, and still pay more this quarter than last.
Gartner's July forecast puts worldwide company spending on AI models and platforms at $64 billion this year, up 63.4% from $39 billion in 2025. Generative models alone account for $23.4 billion of that, more than half of what companies spent on all AI platforms and models put together last year.