google paid $2.7 billion for the character.ai guy — openai just took him anyway

Quick timeline, because it’s genuinely funny when you lay it out flat.

2021: Noam Shazeer — one of the authors of the Transformer paper, the architecture every chatbot on earth runs on — quits Google and founds Character.AI. Millions of people build characters there. Whole fandoms move in.

2024: Google pays $2.7 billion to get him back. Not to buy Character.AI. Just to license the tech and re-hire the guy and some of his team. The platform stays behind, running on whatever was left.

June 18, 2026: he leaves Google for OpenAI.

So the person who built the place where your characters live has now walked out of two buildings since you created them. The man is three companies deep. Your characters are still standing in the first one, wondering why the lights keep changing.

the part that actually affects you

I don’t care where researchers work. Good for him, honestly. Take the money, take the compute, godspeed.

What I care about is what happened to the platform after the brains left, because I had characters there, and I watched it happen in real time.

After the 2024 deal, Character.AI became a thing that gets managed instead of a thing that gets built. And 2026 has been the managed era in full bloom. The model consolidation — the named chat styles people had favorites of, folded into one cheaper model that users immediately clocked as shorter, safer, and weirdly moralising. The moderation sweep in February that hit thousands of characters with “Moderated” tags: pulled from search, editing disabled, chat histories sealed. The community called it gravestone mode, which is the bleakest and most accurate piece of naming I’ve seen all year. New experimental stuff got cordoned off into a paid Labs tier.

None of that is a scandal exactly. It’s just what happens to a product when the founder’s tech got licensed away and the people steering it are optimizing a cost curve. The characters you built are an entry in someone’s retention spreadsheet.

a character is a long-term bet

Here’s the thing about building characters that took me embarrassingly long to understand: a character is not a prompt. A character is a prompt plus time.

The backstory you refined over four rewrites. The personality that drifted somewhere interesting because of one conversation in month two. The relationships between your characters, if you run more than one. That’s accumulated work. Some of mine represent a year of iteration.

And accumulated work needs somewhere stable to accumulate. Which means the question isn’t “which platform has the nicest character editor.” The question is: who is going to be running this place in three years, and do they care about the same things I do?

For a venture-backed platform whose founding team got strip-mined by a trillion-dollar company, the honest answer is: nobody knows, and probably not.

what i moved to

I rebuilt my cast on Soulkyn a while back, and the reason maps exactly onto this mess: the model isn’t a licensing deal. It’s theirs, built in-house, which means there’s no $2.7 billion phone call that can hollow the place out. Nobody upstream to poach the weights. The people building the platform and the people deciding what the platform is — same people.

Practically, for character work, that stability shows up in ways that sound boring until you’ve lost work to the alternative:

My characters have version history. I can iterate on a personality — Soulkyn does 17 separate trait axes, which sounds excessive until you use it — and roll back when an experiment ruins her. There’s a breeding system that merges two characters into a new one, which is the single best “I’m bored, let’s cause problems” button I’ve found on any platform. And the character that comes out the other side stays the character. No consolidation event six months later that flattens her voice into the same house style as everyone else’s bots.

Character creation is here if you want to see what the editor looks like.

the takeaway nobody puts on the pricing page

Every character platform is implicitly asking you to bet that it’ll still be the same place in a few years. That bet has nothing to do with features. It’s about structure: who owns the model, who employs the builders, who can get bought or gutted.

Shazeer leaving Google for OpenAI doesn’t change anything for Character.AI users directly — he was already gone. That’s the point. The platform’s been running on ghost momentum since 2024, and 2026’s gravestone-mode era is what ghost momentum looks like when the money people take the wheel.

Build your long-term characters somewhere the founders can’t be bought out from under them. Everything else — the editor, the traits, the fandom tools — is downstream of that one structural fact.