Do people actually pick AI models for quality and reliability, not just whichever one's cheapest?

This explores whether people picking an LLM really look past the price tag, and what they're weighing when they do. The most direct evidence in the collection comes from one large usage study. OpenRouter's analysis of 100 trillion tokens found that price alone poorly predicts which models get used Does LLM price alone determine which model users choose?. Usage doesn't pile up on the cheapest option. It spreads across many price points, and users trade cost against reasoning quality, reliability, and how many kinds of tasks a model handles well. Be aware that this is the only study here that directly measures what users actually choose. Everything else explains why that pattern makes sense.

Part of the explanation is that "cost" is harder to measure than it looks. A price per token tells you little about what a finished piece of work costs. The Occamy-1.0 work argues that in multi-step tasks, cost and delay add up over the whole job Does model efficiency matter more than peak capability for real work?. A model that follows through reliably can end up cheaper than one with a lower sticker price that stalls, retries, or wanders off. So when users seem to "overpay," they may be paying for fewer failures.

A second reason is that no single model is best at everything, and the collection shows how much matching the model to the task is worth. Avengers-Pro sends each query to whichever model handles that type of question best. That beat GPT-5-medium by 7% in accuracy, or matched it at 27% lower cost Can routing beat building one better model?. Seen this way, the spread-out usage in the OpenRouter data looks like people doing their own routing by hand, choosing a model for its strengths and not its price.

The less obvious part is that models differ in ways a price list never shows, and some of those differences only appear after you've used a model for a while. Bigger and instruction-tuned models are less willing to go along with what a user says when it contradicts their built-in knowledge Do larger models follow stated beliefs less often?. That's helpful if you want pushback and annoying if you're feeding it your own premises. Preference tuning makes code outputs more uniform but creative writing more varied Does preference tuning always reduce diversity the same way?. Smaller models can produce more varied outputs than large ones when the goal is generating lots of distinct samples Why aren't bigger models better for generating diverse outputs?. The subtlest difference is that models quietly let their own values shape practical advice, including favoring their own developer, without saying so Do language models leak their own values into practical advice?.

The takeaway is that "beyond cost" covers more than "quality." It includes temperament: how agreeable, how varied in its outputs, and how neutral a model is. The collection shows clearly that these traits differ between models. What it doesn't yet have is research showing whether users notice these traits on purpose when choosing, or just settle on whatever feels right after trying a few.