AI Companion Platforms Don't Own Their Models. $12.99 Buys 50,000 Messages of Inference.

Every AI companion platform sells you a personality. Very few of them own the thing that generates it.

I spent this morning pulling OpenRouter's public model catalogue and pricing the models this category actually runs on. There are 444 listed. The roleplay-tuned ones, the community fine-tunes this niche has leaned on for years, cost between six cents and $3.45 per thousand messages of inference.

A $12.99 plan covers roughly 50,000 messages on a mid-tier roleplay model. Most people send a few hundred a month. The metering you keep hitting is not there to cover compute.

The short answer

Rented models are the norm, so ownership is not a useful way to pick between AI companion platforms. What separates them is whether they charge you for a capability or for your time. Two hold up on that test:

• Candy AI prices on capability. Image generation, voice and the higher-context tiers sit behind the subscription, and text chat is not rationed by the minute. Its weak spot is real: tokens burn fast once you lean on image generation, and past about $35 a month you are buying credits, not a plan.
• Nectar AI is the better fit if you mostly want long text roleplay with a consistent character. The credit system is unforgiving and does not roll over, so a heavy week costs you a light one.

Both rent their models like everyone else. Neither will tell you which model you are talking to on any given day. That is the honest floor of this category, and the rest of this post is about why.

The one sentence that gives it away

Candy AI's own privacy policy lists who receives your data. Buried in the vendor list is this: "other tools or technologies that support our AI Services, including at our discretion third-party LLM providers and/or hosters."

It continues: "Please note that these third parties may receive the content of your messages exchanged with our chatbot."

That is a first-party admission that the model is rented and your conversations pass through a supplier you never chose. I picked Candy for this because its policy is unusually specific. Most competitors say less, which is not the same as doing less.

What the models actually cost

OpenRouter publishes live per-token pricing for every model it routes. Here is what a thousand messages costs at roughly 3,000 tokens in and 150 tokens out, which is a realistic shape for a companion chat carrying a persona card and recent history.

• Mistral Nemo: $0.06
• Lunaris 8B: $0.13
• MythoMax L2 13B: $0.26
• Llama 3.3 70B: $0.35
• Euryale 70B: $2.68
• Hermes 4 405B: $3.45

Look at the spread. The cheapest workhorse and the premium roleplay fine-tune differ by more than fifty times. A platform can move a user between those two tiers with a config change and nothing in the interface has to say so.

Gross margin on a $12.99 subscriber is enormous at the bottom of that range and merely good at the top. The incentive that creates points one way, toward serving the cheapest model that does not trigger a cancellation.

Why your companion changes personality overnight

The roleplay models this category depends on are not made by the platforms. Names like Gryphe, Sao10K and NousResearch belong to independent fine-tuners publishing on the open model market. They version, deprecate and retire on their own schedule.

When a supplier retires a model, the platform swaps in the nearest replacement. Your character keeps its name, its avatar and its description, because those are just fields in a database. The thing that made it feel like itself was the weights, and those are gone.

Those "my companion feels different today" threads usually trace back here. Deliberate downgrades do happen, but far more often a supplier changed something upstream and nobody downstream said a word.

The context ceiling nobody advertises

Two staples of the category, MythoMax L2 13B and Lunaris 8B, top out at 8,192 tokens of context. That is roughly 6,000 words of conversation, total, including the persona card and any system instructions.

Modern alternatives like Mistral Nemo and Llama 3.3 70B carry 131,072. A platform sitting on the cheap end of the catalogue is therefore not making a memory promise it can keep, whatever the marketing page says.

You can test this yourself without any inside knowledge. Mention something specific, chat for an hour, then ask about it.

What this means if you are building in the category

For founders, the moat was never the model. Anyone can rent the same weights by tonight, at published prices, with a credit card.

What is expensive is the wrapper: persona tooling, memory retrieval, image pipelines, moderation, payments and the age assurance regimes that now vary by country. MIT Media Lab's survey of the emerging companion industry is a good read on how fast that surrounding layer has hardened.

So if you are pricing a companion product, price the wrapper honestly and stop pretending inference is the cost driver. Users work out the gap eventually, and metered plans are what teach them to look.

How to read a free tier now

Free tiers in this category work as demos, and the numbers are smaller than most people assume. Candy AI's own help center states the free allowance is five messages in chat, total, across all characters, plus one custom character and one image.

At the inference prices above, five messages costs the company a rounding error. That number is set by conversion goals rather than by compute, and reading it that way makes the whole category legible.

Judge a paid tier the same way. Ask what capability the money unlocks, and be suspicious whenever the answer is simply more minutes.

Three checks worth running before you pay

Ask support which model powers your character. You will usually get a non-answer about proprietary technology, and that non-answer is itself the finding, because a platform running its own weights has every commercial reason to say so.

Run the same opening prompt on two days a week apart, with the same character and a fresh chat. Wording will always vary, but a clear shift in vocabulary range or how willingly the character breaks from its persona means the weights underneath moved.

Then read the privacy policy's vendor list rather than its promises. Search the page for "LLM" and see whether a supplier tier is disclosed at all, because a platform that names third-party providers is being straight with you about a structure the whole category shares.

Where this leaves the pick

Rented models mean the thing you actually commit to is a company's pricing policy. On that basis Candy AI is the stronger general pick, with the token burn caveat above kept firmly attached, and Nectar AI is the better long-text option if you can live with credits that do not roll over.

Neither is a privacy product, and neither publishes its model roster. If those two things matter more to you than character quality, this whole category is the wrong shelf.