AI chat pricing: what each message costs and how to charge for it

AI chat pricing starts with one number: what a single reply costs to generate. On current open-weight models it is roughly a tenth of a cent, so text chat can sit inside a $10 to $14 subscription. The margin risk is not the average user. It is long conversations and heavy users, which caps, context limits and credits control.

This guide is for affiliates, media buyers, agencies and creators who run or plan to run an AI companion site. It covers the cost side first, with token math you can check against live price lists, then the five ways sites charge for chat, then a worked margin example. If you want the full picture of subscription tiers, token packs and bundles, read our companion piece on AI companion site pricing. This article stays focused on chat itself: what it costs you per message and how to turn that into a price.

How much does one AI chat message cost to generate?

Language models are billed per token, a chunk of text of roughly three quarters of a word. Every reply has two parts: the input tokens you send (the system prompt, the character's persona, any memory notes and the recent conversation) and the output tokens the model writes back. The input side is usually the bigger number, because the whole conversation context is resent with every message.

A realistic companion reply looks something like this: about 1,500 tokens of system prompt and persona, about 2,500 tokens of memory and recent history, and a reply of about 150 tokens. Call it 4,000 tokens in and 150 out. Here is what that costs on a few models at the prices OpenRouter lists in October 2026:

Two caveats matter. First, these are list prices on one router and they move often, usually downward, so check the current rate before you budget. Second, the closed general-purpose models in the table are there as a price reference. Closed providers' usage policies have generally restricted sexual content and can change at any time, which is why most adult companion operators run open-weight models through hosts that allow it, or host them themselves. Read the acceptable use policy of any provider before you route adult chat through it.

The takeaway is that text is cheap. On a mid-size open model, one dollar buys around a thousand replies. That is the economic reason leading sites can bundle unlimited text into a subscription and charge separately for images, which cost far more per item.

Why do long conversations cost more than short ones?

Because the input grows. A companion that remembers is a companion that resends a lot of text. On message five, the model reads a short history. On message two hundred, it may be reading 8,000 or 12,000 tokens of persona, memory summary and recent turns before it writes a single word. Output length barely changes, but input cost doubles or triples.

Rerun the Llama 3.3 70B math with an 8,000-token context and the cost per reply roughly doubles, from about $0.001 to about $0.0018. That is still small per message. It stops being small when you multiply it by the users who chat the most.

There are three standard ways to keep context cost in check:

• Rolling summaries. Instead of resending the full history, keep the last few turns verbatim and replace older turns with a short running summary. Users still feel remembered; you send a fraction of the tokens.
• A hard context ceiling. Set a maximum input size per call so a marathon session cannot quietly balloon. Pair it with the summary so quality holds.
• Prompt caching where the host supports it. The persona and system prompt are identical on every call, which makes them ideal for caching. Anthropic, for example, bills cached input reads at 0.1 times the base input price, with a cache write costing 1.25 times base and a default five-minute lifetime that refreshes each time it is used. An active chat session hits that window constantly. Caching support and discounts differ by provider and model, so confirm the terms for whatever you run.

What does a user cost you per month?

Per-message cost only becomes useful when you multiply it by real usage. Usage on companion sites is lopsided: most subscribers chat a moderate amount, and a small group chats for hours every day. Model your costs around three profiles rather than one average.

These usage profiles are planning assumptions, not measured averages, so replace them with your own analytics as soon as you have traffic. The pattern holds regardless: a regular subscriber costs pocket change on a mid-size model, while a heavy user on the same model can eat most of a $13.99 monthly fee before you pay for card processing, chargebacks, hosting, moderation or images. The model choice changes the picture by close to ten times. That is why serious operators route most chat to a cheap, good-enough model and reserve bigger models for paid tiers or special moments.

Chat is also only one line in the cost stack. High-risk card processing and chargeback handling usually cost more per subscriber than text inference does. Our guide to fan site payment processors covers that side.

How do AI companion sites charge for chat?

Once you know the cost curve, the pricing options make sense. There are five common structures, and most profitable sites combine two or three of them.

1. Unlimited text inside the subscription

This is the dominant model among the best-known brands. Candy AI's paid plans include unlimited text chat plus a monthly token allowance for media, according to RoboRhythms' breakdown of its plans, which lists $13.99 monthly, $8.99 a month billed quarterly and $3.99 a month billed annually. Character.AI's c.ai+ costs $9.99 a month and sells convenience and quality rather than more messages: no peak-hour queue, access to its newest model and unlimited regenerations, per eesel AI's pricing summary.

Unlimited is easy to sell and easy to understand. It works because the cost per regular user is tiny. It needs a quiet fair-use policy behind it so a handful of extreme users cannot erase the margin.

2. A small free allowance, then a paywall

The free tier is a sample, not a product. Candy AI gives a free account five chat messages in total, one custom character and one image, according to the same RoboRhythms review. That is aggressive, and it works because the paywall arrives at peak curiosity. Looser free tiers grow traffic and word of mouth but carry real inference cost with no revenue attached, so set the allowance deliberately and measure free-to-paid conversion against it.

3. Credits or tokens per message

Some sites meter chat directly: every message spends a credit, and users buy credit packs. Metering ties revenue to cost perfectly, which is why most sites use credits for images and other expensive outputs. Nomi, for example, sells credit packs such as 40 credits for $4.99 and 100 credits for $9.99 alongside its subscription, per HollywoodLife's 2026 price roundup. Metering every text message, though, makes users watch a counter instead of the conversation, and that tends to shorten sessions. Use credits where cost is real and visible.

4. Daily caps or a regenerating allowance

A middle path: free or entry-tier users get a set number of messages per day, or an allowance that refills over time. It controls the cost of free users without a hard wall, and it creates a natural upgrade moment when the cap hits mid-conversation. The risk is frustration if the cap is too low for the tier price.

5. Model or memory tiers

Instead of selling more messages, sell better ones. A premium tier can route to a larger model, keep a longer memory window or allow longer replies. This prices the expensive thing (context and model size) directly, and it matches what heavy users actually want. Character.AI's newest-model access on c.ai+ is a version of this idea.

How much does an AI companion cost users?

AI companion pricing for users usually lands around $10 to $14 a month at list price, with deep discounts for longer plans. Candy AI, the best-known adult brand in the category, lists Premium at $13.99 a month, with unlimited text messages and 100 tokens a month for images and other media, according to its subscription page (checked October 21, 2026).

Two lessons for operators sit in that price list. First, text chat is sold as unlimited because one reply costs about a tenth of a cent to generate, so even a heavy chatter rarely eats the subscription. Second, the expensive outputs, images and other media, are metered with tokens, which is where the margin protection lives. The annual plan at $3.99 a month trades monthly revenue for cash upfront and lower churn. If you set your own prices, copy the structure (unlimited text, metered media, discounted long plans) rather than the exact numbers, and test them against your traffic.

How should you set your chat price?

Do not price chat from cost. Cost sets the floor; value and competition set the price. With text costing about a tenth of a cent per reply, almost any subscription price clears the floor. The useful questions are different.

• What does the market anchor on? The leading brands cluster around $10 to $14 a month list price for unlimited text, with steep discounts for quarterly and annual plans. Pricing far above that needs a clear reason, such as a niche character set or a better model.
• Where does cost actually spike? Images, long memory and heavy users. Put those behind credits, a higher tier or a fair-use ceiling, and keep ordinary chat feeling unlimited.
• How much cash do you need up front? If you buy traffic, annual and quarterly plans recover acquisition cost faster, at the expense of a lower effective monthly price.
• What happens at the cap? Every limit is an upgrade prompt. Design the message users see when they hit it as carefully as your landing page.

Revenue concentrates in the operators who get this right. Appfigures counted 337 revenue-generating companion apps in mid-2025, with the top 10% earning 89% of category revenue, and revenue per download rising from $0.52 in 2024 to $1.18, as reported by TechCrunch. Pricing and packaging are a large part of what separates the top tier from the long tail.

Worked example: chat margin on a 1,000-subscriber site

Here is the math on a hypothetical site with 1,000 paying subscribers at $13.99 a month, unlimited text, running Llama 3.3 70B. The usage mix is an assumption for illustration: 950 regular users and 50 heavy users.

Now apply two levers. Add a rolling summary that holds heavy users to a 4,000-token context, and their cost falls from about $11 to about $5.80 each. Route free-tier and low-stakes messages to a smaller model, and total inference can drop by a further meaningful slice. In this example the first lever alone saves about $260 a month, and the bill falls to roughly 12% of revenue. None of those changes are visible to the regular subscriber.

The bigger lesson sits on the other side of the ledger. A 5% change in conversion or a few more annual plans moves revenue far more than any inference trick. Spend most of your testing effort on price points, plan mix and the upgrade moment, and treat inference as a cost to keep in a sensible band rather than the main event. For a fuller view of what a site can earn across every revenue stream, see our AI companion site earnings breakdown.

Build it yourself or use a white-label platform?

If you build your own stack, every number in this article is yours to manage: model selection, hosting, context engineering, caching, moderation, payment processing and price testing. That is a real engineering and operations workload, and it runs before you spend a dollar on traffic. Our AI girlfriend API vs white label guide prices out that route step by step.

On a white-label platform such as WhiteLabelFans, the provider runs the stack, including the models, hosting and billing, and the partner focuses on brand and traffic. Partners keep up to 60% of total site revenue across subscriptions, tips, content unlocks, PPV and upsells, so the inference bill is not a line the partner has to manage. Compare the trade-offs in our guide to the white-label AI companion platform model, or see how a build compares in the Candy AI clone breakdown.

Frequently asked questions

How much does an AI companion cost?

For users, an AI companion subscription typically costs about $10 to $14 a month at list price, less on annual plans. Candy AI, for example, lists Premium at $13.99 a month or $3.99 a month on a 12-month plan (checked October 21, 2026). For operators, the cost to serve one text reply is roughly a tenth of a cent on current open-weight models.

How much does an AI chat message cost to generate?

On current open-weight models, a typical companion reply with about 4,000 tokens of context costs between roughly $0.0001 and $0.001, based on October 2026 OpenRouter list prices for Mistral Nemo and Llama 3.3 70B. Larger closed models cost several times more. Long conversations raise the cost because the history is resent as input with every message.

Should an AI companion site charge per message?

Usually not for text. Per-message charging matches cost exactly, but counters make users ration their conversations, which lowers engagement and retention. Most leading sites include unlimited text in the subscription and use credits for images and other expensive outputs. Per-message credits make more sense on a pay-as-you-go tier for users who will not subscribe.

Why do heavy users cost so much more?

Two effects stack. They send many more messages, and their conversations are longer, so each message carries a bigger context. A user sending 200 messages a day with an 8,000-token context can cost around $11 a month on a mid-size open model, compared with about $1.44 for a regular user. Rolling summaries and context ceilings bring that down sharply.

What is a fair-use policy and do I need one?

A fair-use policy is a published limit that applies to unlimited plans, such as a daily message ceiling or a maximum context length, that only extreme users ever reach. You need one if you sell unlimited text. It protects your margin against automated or marathon use without changing the experience for ordinary subscribers, and it gives you a clean upgrade path.

Do I pay model costs on a white-label AI companion platform?

That depends on the provider, so read the agreement. On WhiteLabelFans, the platform runs the models, hosting and billing, and partners keep up to 60% of total site revenue. Terms differ between providers, for example setup fees or usage-based charges, and any of those changes the margin math in this article.

Sources

• OpenRouter: Llama 3.3 70B Instruct pricing
• OpenRouter: Mistral Nemo pricing
• Anthropic: Prompt caching pricing and cache lifetime
• TechCrunch: AI companion apps on track to pull in $120M in 2025
• Candy AI: Premium subscription plans
• RoboRhythms: Is Candy AI free and what its 5 message limit means
• eesel AI: Character.AI pricing and what c.ai+ includes
• HollywoodLife: How much does an AI girlfriend cost in 2026?