Model Hopping

Programs that write texts or answer questions are now offered by many different companies. They are called things like ChatGPT, Gemini, or Claude. Model hopping means: users don’t stick with one of these programs, but constantly jump back and forth between them. Today someone uses one, in two months the next, because it has become better or cheaper. The word comes from the English “to hop”, meaning to jump. This refers both to individual users and to companies switching their software to a different provider.

Why providers lose their customers so easily

In many industries, customers stay out of habit. Someone who owns a phone from one manufacturer often buys from them again because photos, apps, and accessories are compatible. Experts call such barriers switching costs. With AI programs, these barriers are astonishingly low. You simply type your question into a different window.

For the companies behind these programs, this is a serious problem. They spend billions on data centers and training. Nevertheless, a single better competing product can cause user numbers to collapse within weeks. This is precisely why analysts often speak of AI providers lacking moats, meaning a lack of lasting advantage.

Model hopping also has consequences for pricing. Because customers can switch away immediately, providers undercut each other. The cost for a given amount of computing power has dropped sharply in recent years. Users benefit from this, while providers' profit margins suffer.

What makes switching technically so easy

Developers usually don’t build AI in themselves but call it via an interface. An interface is a fixed address on the internet to which you send a question and from which an answer comes back. Most providers have structured these addresses very similarly. Switching then often just means: swapping the address and access key in the program code.

Many companies go a step further. They deploy an intermediate layer called a router or gateway. This layer decides for each request which provider handles it. Simple questions go to a cheap model, difficult ones to an expensive one. Model hopping then happens automatically, hundreds of times per minute.

Switching is nonetheless not entirely free. Every program reacts somewhat differently to the same instruction. That’s why the phrasings used to steer the program often have to be retested. A typical misconception is confusing model hopping with switching cloud providers. Anyone who moves all their data elsewhere faces a much bigger move.

Model hopping in news and products

The term appears in business news whenever a new model is released. Media then report on how quickly developers migrate to the new provider. Such reports sometimes even move stock prices, for instance at companies that are heavily dependent on a single AI partner.

In everyday life, one encounters this principle in tools for programming or writing. There is often a selection menu with several model names. You choose a fast model for a quick summary and a powerful one for an analysis. Even affordable chat apps that bundle several providers into one subscription thrive on this idea.

Providers therefore try to make hopping less attractive. They integrate memory for personal preferences, link their AI to email and calendar, or offer discounts for long-term contracts. The more of a user’s own data resides in one system, the more expensive it becomes to leave. Whether this is enough remains an open question: so far, model hopping is considered one of the strongest arguments that no single provider will permanently stay ahead in the AI market.