The Cheap AI Model Cost $14 to Fail a Job the Pricy One Did for $1
Each extra step uses billable tokens, and a mismatch between the model and the job can mean paying for failure.

WSJ What’s News · Why Businesses Can’t Budget for AIAsking a cheaper artificial intelligence model to do a hard job can be like hiring the lowest bidder to fix your plumbing, and then paying by the hour while he floods the kitchen.
Stephanie Stam, a visual editor at The Wall Street Journal, ran the same request through two Google models: to study a single still image taken from a YouTube video. Gemini 3.1 Pro, built for complex programming work, finished in 85 steps for about $1. Gemini 3 Flash, built for quick questions, took 952 steps, failed, and rang up about $14.
The reason is how this newer kind of AI, called agentic AI, works. It does not just answer. It thinks, decides, acts, and then thinks again, back and forth. Stam said each of those little decisions uses tokens, the units of work that show up on a company bill. More steps mean more tokens, and the meter runs whether the answer is right or not.
You can watch the difference in the play by play. Pro downloaded the video right away and did the analysis. Flash immediately hit bot detection, finally downloaded the video several steps later, asked another chatbot for the answer, and tried the search engine Bing several times. Along the way it logged its growing frustration with itself. The business would still pay the $14.
That wrong match of model to task is common. A study this spring from researchers at Stanford, Carnegie Mellon, UC Berkeley, and Microsoft concluded that budget models produced much larger token bills than premium models in about a third of comparisons, likely because people used them for jobs they were not built to handle.
Questions people ask
What exactly is a token?
A token is the small chunk of text the model reads, writes, or reasons over. Providers count those chunks and charge a price per million, so a long, rambling attempt costs more than a short, direct one, even if the short one uses a model with a higher listed price.
Do companies really have to pay when the AI fails?
Yes. Most business AI is billed for use, not for success. Every attempt, search, and self correction counts, so a model that tries 952 times and quits still generates a bill. That is why failed runs can be the most expensive ones.
What makes agentic AI different from a regular chatbot?
Agentic AI is software that can take a series of actions on its own to complete a task, such as downloading a file, running code, or searching the web. Each action involves more model thinking, which is why costs can spiral when the agent gets stuck.

