A text assistant in your app can cost a few dollars a month to run and a voice one over a thousand, so the caps belong in the build, not after the first bill.
A five minute voice conversation with an AI agent costs $0.40 in speech alone at ElevenLabs' published rate of $0.08 a minute. The same conversation typed into a chat box costs well under a cent on a small model at Anthropic's list prices. That gap is the decision a Lebanese founder makes when the app quote says AI assistant.
We get this question from founders whose app includes a concierge, a support bot or a voice agent. The build is a one-time number on our mobile app development page. The AI bill is monthly, it grows with every user, and nobody quotes it unless you ask. Here is how to estimate it before launch.

How much does it cost to use an AI chatbot?
Inside your own app, a text conversation costs under a cent on a small model and about six cents on a large one. That is at Anthropic's list prices and our usage assumptions. Platform subscriptions are a different product: Tidio's 2026 survey lists hosted chatbots that bill per seat or per resolved question.
The table uses our working assumptions for a typical app assistant: 3,000 conversations a month, each about six turns, roughly 20,000 tokens read and 2,000 tokens written per conversation once the instructions and history are counted. The prices are the published per million token rates on Anthropic's pricing page and the per minute agent rate on ElevenLabs' API pricing.
| Assistant type | Price used | Cost per conversation | 3,000 conversations a month |
|---|---|---|---|
| Text, small model (Haiku 5.5) | $0.10 in, $0.50 out per million tokens | $0.003 | $9 |
| Text, large model (Sonnet 5.5) | $2 in, $10 out per million tokens | $0.06 | $180 |
| Voice agent, 5 minutes, small model | $0.08 a minute plus the text cost | $0.403 | $1,209 |
Source: Anthropic pricing, ElevenLabs API pricing, read 2026-11-14; usage figures are Voxire working assumptions
On the small model, voice costs about 134 times what text does per conversation. Against the large text model the gap shrinks to under seven times. Either way, voice is the line that decides the bill.

What makes voice different from text in an app?
Voice is billed by the minute while text is billed by the word, so a slow, friendly call costs money every second. At ElevenLabs' agent rate of $0.08 a minute, a caller who thinks out loud for ten minutes costs $0.80 before the model says anything useful.
Text has the opposite shape. A user who types three short questions and leaves costs almost nothing. That is why we start most Lebanese apps with a text assistant and add voice only where the user cannot type: a driver, a caller on a landline, an older customer who prefers to talk.
If you do add voice, it needs a hard ceiling per user and per day. Without one, a single looping session or a curious tester can spend a week of budget in an afternoon.
Does Arabic use more tokens than English?
Yes, the same text usually takes more tokens in Arabic than in English, so the same chat costs more. Research on tokenizer fairness across languages found that one sentence can need up to 15 times as many tokens in one language as in another.
For a Lebanese app that means testing the cost in the languages your users actually write. Many will mix Arabic, English and Arabizi in one message. We measure token counts on real sample conversations in all three before we set a budget, instead of assuming the English figure holds.
It also argues for a short system prompt. The instructions are sent with every turn, so a long Arabic instruction block is paid for again and again. Anthropic's prompt caching cuts the repeated part to $0.01 per million tokens on the small model, a tenth of the normal input price.
What did we build into a property app's AI concierge?
We build an AI concierge for a property listings platform that answers buyers in chat and voice, and three guards came out of it. Each one exists because of a real failure in testing, not a checklist, and each one caps cost as much as it improves the experience.
- A voice budget that ends gracefully. When the session's voice allowance is spent, the panel says so and moves the buyer to text instead of failing silently.
- Retry that resends one turn. If a reply fails, Try again resends only that turn, not the whole conversation, so a flaky connection does not double the bill.
- One lead per conversation. The concierge confirms the buyer's brief, streams its answer and routes a single lead to the right agent, then updates that lead instead of creating duplicates.
None of these show up in a demo. They show up in the third month, on the invoice and in the sales team's inbox. We scope them into every AI feature we build, as part of the fixed price on our app development service.
Can I buy my own AI bot?
Yes, you can subscribe to a hosted chatbot or build one into your own app, and the choice is about control more than price. Tidio's pricing review puts a custom build at $5,000 to $250,000 upfront plus 15% to 20% of that every year.
A hosted bot is faster to start and fine for answering questions on a website. It is weaker when the assistant must read your own listings, orders or bookings, speak your users' mix of languages, or hand a lead to your team in a specific way. Our guide to what an AI chatbot on your website can safely say covers the hosted side.
If the assistant is the product, build it. An MVP app costs $8,000 to $12,000 in our published price list, and the AI feature is scoped inside that number, not billed as an add-on.
Which model should a Lebanese startup start with?
Start on the smallest model that answers your test conversations correctly, and move up only for the turns that need it. At Anthropic's list prices the large model costs 20 times the small one per input token, so routing matters more than the model name.
In practice we send routine turns, such as greetings, opening hours or listing details, to the small model. Harder turns, such as comparing three properties or explaining a payment plan, go to a larger one. The user never sees the switch.
Budget for the running cost the same way you budget for hosting. Our app maintenance cost guide lists the other monthly lines, and our app cost guide for Lebanon covers the build.
What should you actually do this week?
Ask your developer for a running cost estimate per conversation, in text and in voice, before you sign. A good answer names the model, the token counts and the caps. The ElevenLabs and Anthropic price pages give you everything needed to check the arithmetic.
- Write ten real conversations your users would have, in the languages they use.
- Count tokens for each and multiply by the published rates.
- Decide where voice is truly needed, and cap it per user per day.
- Ask how failed replies are retried and how duplicate leads are prevented.
- Put a monthly spending limit on the AI account before launch.
Then launch with the cap low and raise it when you see real usage. Lowering a bill after it arrives is much harder.
Sources
- Anthropic, API pricing, read 2026-11-14
- ElevenLabs, API pricing, read 2026-11-14
- Tidio, chatbot pricing 2026, read 2026-11-14
- Petrov and others, Language Model Tokenizers Introduce Unfairness Between Languages, read 2026-11-14
- Elfsight, how much does a chatbot cost, read 2026-11-14
Need your assistant scoped before you build?
Send us the assistant you have in mind and the languages your users write in. We will estimate the cost per conversation in text and voice, and tell you which caps the build needs.
→ Mobile App Development → Get a Free Quote → Chat on WhatsApp
Enjoying this article?
Enter your email and get a clean, formatted PDF of this article: free, no spam.
Voxire
Mobile App Development
React Native and native iOS and Android for MENA. RTL support done properly, regional payment integration, App Store and Play Store submission, and the architecture decisions that keep a small app maintainable.
Learn more



