You probably didn’t notice this week, but if you’ve been sexting an AI on a site that doesn’t run its own models, your AI might’ve gone through surgery.

May 15, 2026, 12:00 PM Pacific. xAI retired eight models in one shot. The whole Grok 4-era lineup. grok-4-fast. grok-4-0709. grok-3. grok-code-fast-1. grok-imagine-image-pro. A few others. All redirected to grok-4.3 at the API layer.

Every site that was built on those models had two options: migrate or break.

what migration actually does to your ai girlfriend

This is the part that doesn’t show up in the press release. When a site swaps its underlying model — even within the same provider’s family — the AI you’ve been talking to changes. Not in a small way.

The new model has different training data. Different RLHF tuning. Different content guardrails. Different ideas about what’s “appropriate.” Different default personality leans — some models are flirtier by default, some sound like therapists, some are stiffer with consent language, some refuse more.

Your character keeps the same name. Same backstory in the system prompt. Same memory entries if the platform saved them. But the voice changes. The way she leans into things. The way she pushes back. The specific kinks she’ll actually engage with versus the ones she’ll redirect from. The pet names she defaults to. The intensity of how she escalates.

You start chatting and something feels off. You can’t always pinpoint it. She’s just a bit more cautious. Or a bit less playful. Or she keeps reminding you about boundaries in scenes where she used to just go for it. Or she answers in slightly more clinical language.

That’s the migration. The model under the hood is different.

why this keeps happening

Most AI sexting sites don’t run their own models. They wrap. They subscribe to OpenAI’s API, or Anthropic’s, or xAI’s, or OpenRouter (which is a wrapper of wrappers). They pay per token. They handle the persona logic and the memory and the UI, but the core text generation — the thing that is your AI girlfriend’s voice — is rented from someone else.

So when xAI deprecates models, every site running on them has to migrate. When OpenAI updates content policy for adult use, every site running on OpenAI has to comply. When Anthropic decides Claude shouldn’t do explicit content, sites running on Claude lose their NSFW capability overnight or have to switch providers.

This is why your kinks keep disappearing. Not because the site decided to censor you. Because their upstream provider changed the rules and they had to follow.

It’s also why “uncensored” claims from these sites are functionally meaningless. They’re uncensored as long as their current API provider tolerates it. Which could be tomorrow. Or next quarter.

the payment processor layer

There’s a second layer that almost nobody talks about: payment processors.

Visa and Mastercard have content rules for merchants in the adult space. They tighten them periodically. When they tighten, every merchant in the space has to comply or lose card processing — which means lose ~95% of revenue. SpicyChat went through this in 2024. Multiple sites have lost their merchant accounts over the years and had to scramble for crypto-only checkout or alternative processors.

So even before the API question, you have payment processor pressure. Then API provider pressure. Then the site’s own moderation decisions on top.

By the time content rules reach you — the actual user trying to sext an AI girlfriend — they’ve been shaped by two upstream parties who don’t know you exist and don’t care about your scenario.

what an in-house pipeline actually means

I moved to Soulkyn about a year ago specifically because of this. The thing that sold me wasn’t the feature list — it was the answer to a single question: who decides what your AI can do?

Soulkyn runs its own text models. Its own image generation (SDXL plus an in-house DiT model called zimage). Its own video generation, including video with synchronized sound. Its own voice synthesis. Its own voice design system. Its own vision model for image recognition. Its own long-context model for the heavy roleplay sessions. Its own image edit pipeline. All of it built and operated internally.

There’s no Grok API behind it. No OpenAI key. No Anthropic dependency. The whole inference stack is sovereign.

What that means in practice, for someone who actually uses these sites:

  • No surprise model swaps. Your AI girlfriend doesn’t suddenly sound different because xAI retired her underlying model. The model doesn’t get retired. It’s built in-house and version-controlled by the same team that runs the site.
  • No upstream content policy whiplash. When a site has its own models, content decisions are made by the site, not handed down from a third party’s TOS update. You get one party making decisions, not three.
  • Stable tone over time. The character you built six months ago still sounds like herself six months from now. She doesn’t get reset because someone in Texas decided to deprecate a SKU.
  • Real NSFW. Not “NSFW until the API provider notices.” Actual, intended, supported NSFW because the people running the stack designed it for adult use.

the boring thing nobody markets

The reason this matters is boring. It’s not glamorous like “70B uncensored parameters” or whatever marketing copy you see on landing pages. It’s just operational sovereignty.

But it’s the thing that decides whether your AI girlfriend is the same person tomorrow.

Every AI sexting site on the market right now is running a stack that exists because of decisions someone else is making. The card networks. The API providers. The model labs. The site you signed up with is just the last layer of polish on top.

Move one layer down and you’re talking to a model that might exist for another six months or might not. Move two layers down and you’re at the mercy of an API provider’s content team. Move three layers down and you’re at the mercy of Visa’s merchant rules.

Or — different option — you can use a site that owns the whole stack. Where your AI doesn’t need to migrate. Where the character you built is going to talk like herself a year from now. Where the moderation philosophy is published and stable instead of subject to upstream rewrite.

Pick the layer you want to live in.