The state of AI companions in 2026

AI companions stopped being a curiosity sometime in the last two years. What was a handful of apps in 2023 is now a crowded, fast-moving market: one widely-cited comparison counts over 128 companion apps in active distribution, up from 16 three years earlier, and another market write-up puts the number of active, revenue-generating apps even higher, at 337 worldwide. MIT Technology Review named AI companions one of its 10 Breakthrough Technologies of 2026.

And yet most of the public conversation about them is bad. It splits into two unhelpful camps: breathless hype ("your new AI best friend!") and reflexive dystopia ("the loneliness machines"). Neither tells you how the space actually works — who the players are, what choices separate them, and what is genuinely changing underneath.

This post is an attempt at the missing thing: a map. Not a ranking, not a recommendation — a description of the territory as it stands in 2026, written for people who want to understand it rather than be sold on it.

How we got here

The category predates the current AI boom. Replika launched in 2017, years before large language models made conversation cheap and fluent, and for a long time it was nearly synonymous with "AI companion." The modern explosion came after ChatGPT: once capable language models were available behind an API, the barrier to building a plausible companion collapsed, and dozens of products rushed in.

One event shaped the market more than any other. In early 2023, Replika abruptly removed its erotic-roleplay features, and a wave of disaffected heavy users went looking for alternatives. Character.AI, Nomi, Kindroid, Chai, and others each absorbed a slice of that migration. The lesson the market took from it still drives product decisions today: the most engaged users are also the most sensitive to sudden changes in what their companion is allowed to be.

The map: how to tell the players apart

The temptation is to rank companion apps from best to worst. That is the wrong frame — it produces a list that is out of date in a month and tells you nothing structural. The more useful question is what choice each product made, because the same handful of decisions explains almost every difference you will notice.

By relationship model

The clearest dividing line is what kind of relationship the product is built around.

Companionship and emotional support. Replika is the archetype: built around one ongoing relationship, marketed for emotional support and meaningful conversation.
Character and roleplay. Character.AI sits at the other end — less a single companion than a platform of millions of community-created characters, closer to interactive fiction than to a relationship.
Evolving companions. Nomi and Kindroid emphasize a companion that develops over time. Nomi's draw is its memory and a relationship-progression system where the companion evolves new traits based on your history; Kindroid leans into deep customization and editable memory for people who want to author a character in detail.

These categories blur — most apps do a bit of everything — but the center of gravity is usually obvious, and it tells you who a product is really for.

By what makes it "feel alive": memory

If there is one axis users themselves fixate on, it is memory. The difference between a companion that recalls last week's conversation and one with "goldfish memory" is the difference between something that feels like a relationship and something that feels like a series of disconnected chats. Nomi is repeatedly singled out as the memory leader in the category, and it is no accident that memory and retention tend to be discussed in the same breath.

By business model

How a companion makes money shapes what it becomes. Pricing clusters in a familiar band — paid tiers run roughly $7 to $20 per month, with annual plans bringing the effective price down (Replika around $5.83/mo annual, Nomi around $8.25/mo annual). A few, like Pi, lean on a completely free model.

The revenue is real but smaller than the hype suggests: one breakdown estimates the companion app market crossed around $120 million in revenue — a meaningful number, but tiny next to the platform valuations being discussed, and next to longer-range projections that put the global market near $49 billion in 2026 growing toward the hundreds of billions by 2035. Treat the far-future projections with caution; the near-term revenue figure is the more honest measure of where the money is today.

By positioning: engagement vs. wellbeing

The last axis is the most consequential, and it is where the category is visibly splitting. Most platforms are, by design, optimized to keep you inside the app — longer sessions, more notifications, deeper attachment. As one comparison bluntly puts it, the business models of the major players "depend on users spending more time inside the platform," an incentive that "does not align with user wellness." A smaller set of newer entrants is trying to invert that — defining success as what a user understands or does outside the app rather than time spent in it. Whether that is genuine or merely a marketing posture is one of the open questions of 2026.

What is actually changing in 2026

Three forces are reshaping the category right now, and none of them existed at this intensity a year ago.

Regulation arrived. This is the biggest structural shift. New York's AI Companion Models law took effect November 5, 2025, and California's SB 243 followed on January 1, 2026, making California the first state to require safety protocols and hold operators legally accountable. The requirements are concrete: companion bots must disclose that they are not human, run protocols to detect self-harm and surface crisis resources, and — for minors — repeat the "this is AI" reminder every three hours. SB 243 even carries a private right of action with damages up to $1,000 per violation. Idaho, Oregon, and Washington have added their own rules, and a Kentucky lawsuit signals that enforcement is not theoretical. The trigger for much of this was tragic: wrongful-death suits tied to teenagers, some of which were settled in early 2026.

Safety and wellbeing became a real research question, not a talking point. An Aalto University study presented at CHI 2026 followed roughly 2,000 companion users — mostly Replika — over two years, and reported that the heaviest-use group showed more signals of loneliness, depression, and suicidal ideation over time than comparison groups. That finding deserves careful reading rather than panic — heavy use and distress can be correlated in either direction — but it has moved the wellbeing conversation from rhetoric into data, and the products know it.

Consolidation is starting. With hundreds of apps in the market, the shakeout is beginning. The commentary suggests the platforms with real retention and a clear revenue model will absorb compliance costs and survive, while thinly-monetized apps with no retention are exposed to the next regulation wave or the next jump in model costs.

The open questions

A map shows you the terrain; it does not tell you where the roads will go. The questions I find most worth watching — and that I want to look at more closely in the posts that follow — are these:

Memory. Why does it matter so much to users, and what actually separates a companion that feels continuous from one that does not?
Privacy. These apps hold some of the most intimate data people produce. What really happens to it?
Safety and duty of care. What does responsible design look like for a product people get attached to — especially given the new legal floor?
Wellbeing vs. engagement. Is a companion that genuinely points you outward even possible as a business, or does the economics always pull the other way?

That is the state of AI companions in 2026: a large, crowded, newly-regulated category splitting along a few clear lines, with the hardest questions — about memory, privacy, and whether any of this is good for the people using it — still wide open. The next few posts take them one at a time.