I checked what 10 AI companion apps disclose about their models. Most say nothing.

If you build with LLMs you are used to knowing what you are calling: a model name, a context window, a price per million tokens. Consumer AI companion apps are the opposite. Most of them describe the product as "advanced AI" and leave it there.

I went through the public docs, help centres and pricing pages of ten of the bigger apps to see what they actually publish. Here is what I found, and why it matters if you are evaluating one.

What "disclosure" means here

I looked for four things on first-party pages:

1. Named chat models (or at least named tiers with stated sizes)
2. Context or memory limits in tokens or characters
3. Named image or video engines
4. A price per action where the feature is metered

The open ones

SpicyChat is the most open about chat. Its docs list more than a dozen open-weight roleplay models by name, tie each to a subscription tier and state context sizes. You can see the list in its profile.

Kindroid runs its own versioned model line and publishes memory limits per tier, which is rare. Profile here.

Janitor AI documents its in-house model and lets you bring your own API key, so in practice you can choose the model yourself. Profile.

Nomi names its own models and keeps a public update log. Profile.

The quiet ones

The largest consumer brands tend to be the quietest. Character.AI exposes "chat styles" but not the base models behind them. Replika's only statement is that it combines its own and third-party LLMs. Several adult-focused apps name nothing at all.

Why you should care

• Context size predicts memory. An 8K context app forgets what you said an hour ago no matter how good the marketing is. The memory ranking sorts apps by what they publish and how they behave.
• Model size predicts roleplay quality, up to a point. An 8B model and a 70B model do not write the same scene. See the roleplay ranking.
• Metered features need a price list. If an app will not tell you what an image or a clip costs, the subscription price is not the real price.

A comparison worth reading

The three memory-first apps get compared head to head here, including how each one stores long-term facts: Nomi vs Kindroid vs Replika.

Where the data came from

I used the per-app profiles on AI Companion Radar, which cite the docs and mark anything second-hand as reported. Their scoring rubric is public: how they score.

If you know of an app that publishes more than the four above, tell me in the comments. I would like the list to be longer.