Most consumer software gets cheaper to serve as it grows. AI companions do not work that way. Every message has a marginal cost, every image has a larger one, and every minute of real-time voice larger still. For operators, understanding that cost profile — and designing pricing around it — is the difference between a business that scales and one that subsidises its most loyal users into insolvency.
Cost per active user
The basic unit is simple to describe and surprisingly hard to measure: what does it cost to serve one active user for a month?
Text inference cost is driven by tokens — both the user's message and everything sent to the model alongside it. In companion apps, that "everything" is substantial: a system prompt defining the persona, retrieved memories, a summary of past conversations and the recent message history. A short user message can easily be wrapped in a much larger context. The context, not the conversation, is often the dominant cost.
Add media and the picture changes again. Image generation, text-to-speech, real-time voice and video each carry costs that can dwarf text. A product that looks cheap to run on chat alone can become expensive once its most popular features are media-heavy.
The heavy-user problem
Usage in companion apps is highly skewed. A minority of users send a very large share of messages, and they tend to be the most engaged, most loyal and most likely to pay.
For illustration, consider two subscribers on the same flat monthly plan. One sends a handful of messages a day; the other sends 200 messages a day, requests voice replies and generates images most evenings. They pay the same price, but the second may cost many times more to serve — potentially more than the subscription itself. Flat pricing turns your best customers into your biggest liabilities.
Operators typically respond in a few ways: soft caps or "fair use" limits, credits for expensive actions, tiered plans for heavy users, and engineering work to reduce the cost per message. Most mature products use a combination.
Model routing
Not every message needs the most capable model. A greeting, a short reaction or a simple follow-up can be handled well by a smaller, cheaper model, while a complex emotional conversation, a long roleplay scene or a memory-heavy exchange may benefit from a larger one.
Model routing — deciding per message, or per session, which model should respond — has become a core competency. Routing can be based on message complexity, user tier, conversation type or cost budget. Done well, it can cut average cost substantially with little visible change in quality. Done badly, it produces jarring shifts in personality and tone that users notice immediately.
Small models versus frontier models
The gap between small, open-weight models and frontier models has narrowed for many companion use cases. Fine-tuned smaller models can be very good at maintaining persona, tone and style — arguably the qualities companion users care about most — and they can be self-hosted, which gives operators more control over cost and content policy.
Frontier models still tend to lead on reasoning, long-context coherence and nuanced emotional handling. They also come with usage policies that may not fit every companion product, particularly adult ones.
Trade-off: frontier models buy quality and save engineering effort at a higher and less controllable cost; small models buy control and margin at the price of in-house expertise and hosting.
Caching and context management
Because so much of every request is repeated — the persona prompt, stable identity facts, long-running summaries — caching is one of the most effective cost levers available. Prompt caching from model providers, or equivalent techniques on self-hosted infrastructure, can reduce the cost of the static portion of each request.
Context management matters just as much. Summarising old conversations rather than resending them, retrieving only relevant memories and trimming redundant history all reduce tokens per message. The best memory systems are not just better for users; they are cheaper to run.
Voice and images
Media features deserve their own budget lines. Voice — especially real-time, low-latency conversation — is expensive and scales with minutes, not messages. Image generation costs scale with resolution, model choice and retries. Video is costlier again.
Many operators therefore price media separately, through credits or higher tiers, rather than bundling it into a base subscription. That aligns revenue with cost, although it introduces pricing complexity and draws scrutiny over spending mechanics.
Designing pricing around cost
Sound companion pricing tends to share a few principles:
- Cover the median user comfortably with the base plan, and protect against the tail with limits or credits.
- Price expensive actions explicitly rather than hiding them in an unlimited promise that cannot be kept.
- Make limits feel fair. Users accept caps when they are clear and generous; they resent them when they appear suddenly mid-conversation.
- Track cost per user alongside revenue per user, by cohort and by plan, so pricing decisions are based on margin rather than top-line growth.
What it means for operators
Inference costs are falling, but expectations are rising just as fast — longer memory, richer voice, more images, video avatars. The likely outcome is not that compute stops mattering, but that cost efficiency becomes a quiet competitive advantage. The operators who know their cost per conversation as precisely as their conversion rate will be the ones able to price sustainably — and to survive the users who love them most.
