The Cheaper AI Cost More Per Task: Only 1 in 10 Companies Forecast AI Spending Accurately
In a survey of 396 companies, just 11% forecast their AI spending to within ±10% of the actual figure. Researchers say usage volume and repeated attempts, more than the listed price per token, determine what AI actually costs.
New research suggests AI bills are genuinely hard to predict in advance. Mavvrik, a company that manages AI costs, says some businesses are having to reset their pricing or cancel projects already underway because of unexpected AI expenses. For companies putting AI to work, the issue isn’t only how large the bill is, but how accurately they can budget for it in the first place.
Mavvrik surveyed 396 companies in April and May of 2026. Only 11% came within ±10% of their actual AI spending. That’s 4 percentage points below the 15% recorded in the 2025 survey.
Why a cheaper model can cost more
Even when the advertised unit price is low, finishing a single task can cost more. Researchers from Stanford University, UC Berkeley, Carnegie Mellon University and Microsoft compared eight reasoning models, the kind that work through a thought process before answering. The tests spanned 12 task types, including math, science, coding and agentic work, where the AI handles a job on its own.
Across 336 head-to-head model pairings, 106 cases, roughly 32%, flipped the expected order: the model with the lower listed price per token ended up costing more to complete the task. The researchers calculated prices and usage volumes for input tokens, output tokens, and tokens that are stored and reused. Tokens are the small fragments of text that models process. They also factored in how many attempts it took to finish a task.
The researchers found that models varied widely in how many tokens they used to produce an answer and how often they sent requests to outside tools. Even when the same question was repeated, execution costs differed by as much as 9.7 times. The upshot: picking the cheapest model off a price sheet tells you little about what the work will actually cost.
What to look at when setting a budget
The paper applied prices as of May 1, 2026, and used a single answer-generation configuration per model. The researchers noted that cost comparisons factoring in answer quality need to be treated separately. This comparison also counted money spent on runs that failed to produce a correct answer.
Google said the number of tokens used can vary from request to request. It explained, however, that when a large volume of varied tasks is processed, cost fluctuations average out. The company added that users have access to spending caps they can set and a choice of pricing options.
What this survey and research point to is the gap between AI’s listed price and what it costs to actually finish a job. Companies setting AI adoption budgets need to account not just for unit prices but for real usage volumes and the number of repeat attempts. Worth watching from here: what costs look like once answer quality is weighed in, how much costs fluctuate at high volume, and how effective spending caps are at keeping budgets under control.