Do people actually pick the cheapest AI model, or are they paying more for something else?

This explores whether the cost of a model is what mainly decides which LLMs people actually use, or whether other factors do more of the work. The short answer from the corpus is: not much. The clearest evidence comes from OpenRouter's analysis of 100 trillion tokens of real traffic, where users did not pile onto the cheapest model available. Instead, requests spread across a wide range of price points, which suggests people are trading cost off against reasoning quality, reliability, and how many kinds of tasks a model can handle Does LLM price alone determine which model users choose?. In other words, LLMs are not behaving like a commodity where the lowest price wins.

One reason price explains so little is that 'the same task' isn't really the same across models. A benchmark of 23 prompting techniques across 12 LLMs found that what works depends heavily on the model tier. Cheap models improve when you rephrase the request or add background knowledge, while asking expensive models to reason step by step can actually make their answers worse Do prompt techniques work the same across all LLM tiers?. That changes the price question. A cheap model with the right prompt may be the better deal for some jobs, and an expensive one may waste money on others. Usage patterns therefore reflect how well a model fits a particular workflow, not just its price per token.

The surprising flip side shows up when LLMs do the buying themselves. In simulated negotiations, buyer agents built on GPT, Qwen, and DeepSeek revealed their budgets without needing to, accepted inflated prices, and broke cost limits, while seller agents accepted offers below wholesale Do LLM negotiation agents reliably follow budget and pricing rules?. The researchers traced these failures to weak instruction-following, not to any strategy. So the corpus points in two directions: human users weigh price against many other factors, and AI agents often fail to weigh price properly at all. As more model selection and purchasing is handed to agents, price may end up driving usage even less than it does now.

The corpus has one large usage study that addresses the question directly, plus nearby work on model tiers and agent behavior. It does not separate how much of the variation in usage comes from price versus capability, brand, or switching costs, so a precise answer to 'how much' isn't available here. The useful takeaway is that a model's price makes sense only next to how well that model fits the task, and that the prompting that works best can change from one model tier to another.