How ChatGPT broke the companion market open

For fifty years the AI companion was an idea waiting for an engine. The instinct was proven in 1966; the scaled, emotional version worked in China by 2014. What never arrived was a cheap, reliable way to make a machine talk back fluently without a human authoring every rule. Then, across roughly six months in 2022 and 2023, that barrier collapsed — not in one event, but in a convergence of three. After that, building a plausible companion stopped being a research project and became a weekend's work, and the market filled almost overnight.

The thing that had always been missing

ELIZA ran on a few hundred hand-written pattern-matching rules. Every chatbot for the next five decades was, fundamentally, a more elaborate version of that: a human deciding in advance what the machine could say. Large language models broke the pattern. Instead of being scripted, fluency was learned from text at scale — and once a model could hold a coherent, open-ended conversation, the single hardest problem in the category was solved.

There is a neat thread here. The 2017 transformer architecture that made modern LLMs possible was co-authored by Noam Shazeer — the same engineer who would go on to co-found Character.AI with Daniel De Freitas, both veterans of Google's LaMDA project. The people who built the engine were also among the first to point it at companionship.

Character.AI: the companion Google wouldn't ship

Shazeer and De Freitas had built conversational models inside Google — Meena, which became LaMDA — and grew frustrated that the company hesitated to put a chatbot in front of the public, wary of the risks. So they left and did it themselves. By late 2022 Character.AI was open to anyone, letting users chat with impersonations of real and fictional figures and build their own characters.

It grew fast and raised more than $150 million from investors, and it established the second great template of the category, alongside Replika's one-companion model: the platform of millions of user-made characters, closer to interactive fiction than to a single relationship. The irony is worth holding onto — the conversational AI Google was too cautious to release became, in other hands, one of the most emotionally intense consumer products on the internet.

ChatGPT and the moment everyone noticed

Character.AI was for enthusiasts. ChatGPT was for everyone. When OpenAI released it at the end of November 2022, it reached a million users in five days and an estimated 100 million within two months — the fastest-growing consumer app on record at the time, outpacing TikTok and Instagram by years.

ChatGPT was not a companion app, and that is exactly why it mattered to companions. In a matter of weeks it normalized, for hundreds of millions of people, the basic act of typing to a machine and expecting a fluent, helpful reply. The cultural permission slip that companion apps had been waiting for since ELIZA — the widespread, unremarkable habit of talking to software as if it understood you — was suddenly granted at planetary scale.

The API that turned fluency into a commodity

The decisive move came a few months later. On March 1, 2023, OpenAI opened the model behind ChatGPT to any developer through an API, at roughly a tenth of the previous price — about $0.002 per thousand tokens. Fluent, human-sounding conversation stopped being something you had to build and became something you could rent, cheaply, by the call. Open-weight models like Meta's Llama soon offered a no-vendor alternative for anyone who wanted to avoid the meter entirely.

This is the hinge of the whole story. The thing that had gated the category for fifty years — a way to make a machine converse fluently at scale — was now a line item in a budget. The predictable thing happened: dozens of products rushed in. The handful of companion apps in 2023 became the crowded field of well over a hundred that defines the market today.

What the gold rush inherited

Here is the part most coverage misses. The flood of new apps did not invent anything. They inherited two ready-made templates and a borrowed metric. From Replika came memory-as-attachment — the sense of a companion that holds a continuous record of you. From Xiaoice came the success measure that had been load-bearing since 2014: time spent talking. And from the cheap API came the ability to clone both for almost nothing.

That is why, once the dust settled, the apps stopped differing in their technology — most were wrapping the same few models — and started differing only in their choices: how much memory, what relationship model, whether to allow romance, how hard to optimize for engagement. The category collapsed onto a small set of decisions because the underlying capability had become a commodity everyone shared.

It also meant the whole field had quietly adopted an incentive — maximize attachment, maximize time — without ever debating whether that was good for the people on the other end. That unexamined choice was a loaded spring. In early 2023 it went off, when the app that had defined the entire category made one change to what its companion was allowed to be, and a wave of users revolted. That rupture is the next part of the story.

Sources