How much does an AI co-host cost? Price per message, per call minute and per text

One AI reply costs from about a tenth of a cent to 4.8 cents at the model provider, depending on the model and caching. Supalobby measured about 1 US cent per AI message on a September 2026 sample. A three-minute AI phone call costs 16 to 31 cents raw, and one emoji can triple the price of a text.

Below is the arithmetic, from vendor price lists checked on 26 September 2026: what a guest conversation costs in tokens, what goes into a minute of an AI phone call, why a curly apostrophe triples a text, when WhatsApp is free, and what AI co-host tools charge on top. PLN prices are converted at the NBP rate of 25 September 2026.

What one message costs in tokens

A language model bills per token, a chunk of text, separately for what it reads (input) and what it writes (output). Prices per million tokens:

Anthropic charges 25 percent extra for the first write to its 5-minute cache; OpenAI does not. Anthropic also notes that Claude models from 4.7 onward split the same text into about 30 percent more tokens than older ones, so comparing price lists alone is not enough.

An agent never reads just the guest's last message. Every reply re-reads its instructions, the property's knowledge base and the conversation so far. Our assumption for the sum: a guest conversation has 6 messages and 6 replies; each reply reads 8,000 tokens, of which 7,000 are the fixed part (instructions and knowledge base) that can be cached; each reply writes 300 tokens. We leave out reasoning tokens, which some models bill as output.

Claude Sonnet 5 without caching: 6 × 8,000 = 48,000 input tokens at $2 per million is $0.096. Output: 6 × 300 = 1,800 tokens at $10 per million is $0.018. Total $0.114 for the conversation, 1.9 cents per reply. With caching, the first reply writes the 7,000 fixed tokens at $2.50 per million and the next five read them at $0.20. Total $0.0545, 0.9 cents per reply.

Two things stand out. Caching roughly halves the bill, because most of what the agent reads never changes. And the choice of model moves the cost thirtyfold. A small model can answer "where do I park"; a request for three nights with a dog and a company invoice benefits from a stronger one.

What a minute of an AI phone call costs

The raw cost of a three-minute call is 16 to 31 cents on a US or UK number. Four parts: the carrier's inbound minute, turning the guest's speech into text, the language model, and the agent's voice.

The assumptions are ours: the agent speaks for half the call, about 1,600 characters in a minute and a half, and each of 8 turns reads 6,000 tokens. With ElevenLabs Multilingual v2 at $0.10 per 1,000 characters, option B rises to about $0.25. We could not tell from ElevenLabs' page whether its per-minute agent price includes the language model, so we add it separately.

Then the number itself. Twilio charges $1.15 a month for a US local number, $3.50 for a UK local number and $2.50 for a UK mobile number. If the agent forwards a call to your own phone, the outbound leg costs $0.0140 a minute in the US and $0.0305 a minute to a UK mobile.

What a text costs, and why one character can triple it

A text is billed per segment, and the segment length depends on the encoding. GSM-7 fits 160 characters in one segment and 153 per segment in a longer message. Any character outside that alphabet, such as an emoji or a curly quotation mark, switches the whole message to UCS-2: 70 characters in one segment, 67 per segment in a longer one. Twilio's own example is a text editor turning a straight quote into a curly one.

This message has 157 characters and a straight apostrophe:

Hi Sarah, it's Anna from Sunny Flat. Door code: 4821#. Check-in from 3 pm, parking behind the building, space 12. Wi-Fi password: Sunny2026. Enjoy your stay.

As written, it is 1 segment. If your phone swaps the apostrophe in "it's" for a curly one, or you add a smiley at the end, it becomes 3 segments.

For automated texts: type them in a plain-text editor, drop the emoji, and keep them under 160 characters.

Does replying on WhatsApp cost anything

Not when the guest wrote first. Since 1 July 2025 Meta charges per delivered template message rather than per conversation. A guest's message opens a 24-hour customer service window in which free-form replies are free, and so are utility templates sent inside it. If you reply within 24 hours to someone who came from an ad or a call-to-action button, a 72-hour window opens in which every message is free.

You pay for templates sent outside the window: a pre-arrival reminder a week after booking, a review request after checkout, an offer for next season. The rate depends on the category and the guest's country code. FormBeep, which mirrors Meta's rate card and last verified it on 10 September 2026, lists per-message rates from 1 July 2026:

Meta publishes the full rate card as a downloadable file; we took these three rows from FormBeep and did not open the file itself.

What AI co-host tools charge

A packaged AI co-host costs from a few dollars to several hundred a month, a multiple of the raw cost. Prices checked 26 September 2026.

The gap between raw cost and price is not all margin. It pays for channel integrations, the knowledge base, handing conversations to a human, failed attempts and retries that also burn tokens, and someone to call when something breaks. ChatLab does not say how many credits one reply uses, so a credit is not the same as a message.

How to estimate your own monthly bill

1. Count guest messages in a normal month. Go through one peak month of Airbnb, Booking.com, WhatsApp and e-mail threads. That is your number of replies.
2. Multiply by the per-reply price of the tool you are considering. With a message pool, check you fit in it. With credits, ask how many one reply uses.
3. Count phone minutes separately. Calls per month times average length. At a few dozen minutes the monthly fee matters; at hundreds, the per-minute rate does.
4. Check whether you send texts or WhatsApp templates outside the 24-hour window. Those are the only messages you pay the carrier or Meta for.
5. Set it against your time. Weigh the tool's price against the hours you now spend replying.

Where Supalobby fits

Supalobby charges per unit, not per token: 25 PLN net on Copilot (40 messages), 49 PLN on Autopilot (100), 99 PLN on Operator (200, announced, not yet purchasable), plus 0.39 PLN for every message beyond the pool. Its measured cost on a sample of 434 conversations from 6 to 13 September 2026 was 0.037 PLN per AI message; that is our cost, not your price. The Voice add-on gives the agent a Polish landline number for 129 PLN, about 29 euro, with 30 minutes, then 2.99 PLN a minute. It does not offer UK or US numbers. Guest texts come only with the Voice + SMS add-on. No commission on bookings. Supalobby launches soon; for now, sign-ups join a waitlist.