We priced a typical customer conversation on a website assistant: three questions, three answers, with the assistant reading your help content each time. On Anthropic’s Claude Sonnet 5.5, a capable mid-range model, the AI costs about 2.9p for the whole conversation. On the small models built for high volume, it costs between about a tenth and a third of a penny.
Buy the same assistant ready-made and the price is per resolved conversation. Intercom charges $0.99, about 74p. Zendesk charges £1.30 for each extra resolution bought in advance, or US$2.00 pay-as-you-go.
Our verdict is that the model is the cheapest line in the budget. Most of the money goes on what surrounds it: the chat window, handing over to a person, reporting, and above all the content the assistant answers from. Budget for those, and treat the model as a rounding error.
What one conversation costs on the AI models’ own price lists
AI models are priced by the token, roughly three-quarters of a word. You pay for everything the model reads (input) and everything it writes (output). In a website chat, the model reads far more than it writes, because every reply means re-reading its instructions, the help content it has found, and the conversation so far.
So we built a realistic example and priced it.
The conversation: three customer messages of about 25 words, and three replies of about 60 words.
What the assistant reads each time: 2,000 words of instructions, about your business and how to behave, plus 1,500 words of help content found for the question, plus the conversation so far.
The conversion: about 1.25 tokens per word. We measured three samples of English prose with OpenAI’s own tokeniser and got between 1.18 and 1.31. Anthropic says its newer models produce “approximately 30% more tokens for the same text”, so we added 30% for Claude.
The prices: the providers’ standard price lists, read on 24 October 2026, in dollars as published. We converted at the European Central Bank’s reference rates for 23 October, which put £1 at about $1.337.
| Model | Size | Cost of one conversation |
|---|---|---|
| OpenAI gpt-6-luna | Small | 0.11p |
| Claude Haiku 5.5 | Small | 0.14p |
| Gemini 3.5 Flash-Lite | Small | 0.35p |
| OpenAI gpt-6.1-sol | Mid-range | 2.19p |
| Gemini 3.1 Pro Preview | Mid-range | 2.23p |
| Claude Sonnet 5.5 | Mid-range | 2.85p |
| Claude Opus 5.5 | Large | 5.70p |
At 1,000 conversations a month, that is between about £1.10 and £3.45 on the small models, and about £28.50 on Claude Sonnet 5.5. Caching the fixed instructions, which all three providers support, cuts each figure by about a quarter.
Your numbers will differ. Longer help articles, longer conversations and models that “think” before answering all cost more. But a conversation would have to grow many times over before the model cost reached the price of a ready-made product.
What a ready-made assistant costs
The three ready-made products below all sell an AI assistant that sits on your website and answers from your help content. Their prices, read on their own pricing pages on 24 October:
| Product | How it charges | Price |
|---|---|---|
| Intercom Fin | Per outcome | $0.99 (about 74p), on any Intercom plan, or with your existing helpdesk subject to a minimum monthly commitment |
| Zendesk AI agents | Per automated resolution, with some included in every plan | Plans from £15 per agent a month, paid yearly. Extra resolutions £1.30 each if committed in advance, or US$2.00 pay-as-you-go |
| Tidio Lyro | Monthly bundle of AI conversations | From $32.50 (about £24.30) a month for 50 conversations, billed annually, which is about 49p each if you use them all. $39 (about £29.20) billed monthly |
Read the definitions closely, because they decide what you pay for. Intercom counts an outcome when a customer confirms their issue is resolved, “or They don’t ask for more help after Fin responds”, or Fin completes a workflow, “including handoffs”. By that definition, a customer who stops replying counts the same as one whose problem was solved. Intercom charges once per conversation, however many questions are answered. Zendesk counts a request “successfully resolved by the AI agent, without any escalation to a human agent”.
Both Intercom and Tidio publish how often their assistants resolve conversations. Intercom says “Fin customers average a 76% resolution rate, across 12,000+ businesses”. Tidio says its customers “average a 75% resolution rate with Lyro”. These are the vendors’ own figures, measured by their own definitions, so treat them as a ceiling for planning rather than a forecast.
What the gap pays for
At 74p against under 3p, the gap looks like profit. It mostly pays for things you would otherwise have to build and run yourself:
The chat window and the inbox. The ready-made products come with the widget on your site, a place for your team to pick up conversations, and the history in one place.
Handing over to a person. When the assistant cannot help, the conversation has to land with someone who can, with the context attached.
Reporting and testing. Seeing which questions it got wrong, and whether editing the content fixed them.
Keeping up with the models. Providers retire models on fixed dates, and a newer one answers differently. With a ready-made product, the vendor chooses the model and handles the move. We covered why that matters in a smarter AI model can make your system worse.
If you build your own
The running costs of a self-built assistant are small at this scale, on the published prices:
Hosting. Cloudflare’s paid Workers plan has a $5 (about £3.74) monthly minimum per account, and includes 10 million requests a month.
Search over your content. One common way for an assistant to find the right help article is a “vector database”. Cloudflare’s own worked example, 25,000 entries searched 50,000 times a month, comes to $0.59 (about 44p) a month before the allowance its paid plan includes. Pinecone’s Builder plan is $20 (about £15) a month flat.
The model. From about £1.10 a month on the smallest model to £28.50 on Claude Sonnet 5.5, at 1,000 conversations, from the table above.
So the running cost of a self-built assistant on Claude Sonnet 5.5, at 1,000 conversations a month, comes to between about £32 and £47 depending on the search service. The real cost is the build and the upkeep: someone to make it, test it, read what it gets wrong, and move it to a new model when the old one is retired. That is a person’s time, every month, and it is the line that decides whether building makes sense.
Model prices move too. Google’s price list, as captured on 23 October, shows Gemini 3.8 Flash at $0.75 per million input tokens until 31 December 2026 and $1.50 from 1 January 2027. A budget set per conversation, and checked against the price list each quarter, will cope with that. A budget that assumes today’s price will not.
The cost on nobody’s price list
Every assistant in this article, bought or built, answers from your content. If your help pages are out of date, it gives out-of-date answers confidently. If your returns policy lives in one person’s head, the assistant will not know it.
The work of writing down what your business knows, in a form an assistant can use, is the same whichever route you take. It can be the largest piece of work, and it decides whether the assistant is any good. A ready-made product at 74p a chat with good content will beat a self-built one at 3p with poor content.
What to do before you buy or build
Count your conversations. Look at your inbox, phone log or live chat for a typical month. The number decides whether 74p a chat is £50 a month or £5,000.
Pull your twenty most common questions and check that your website answers each one clearly. Fix the gaps first. You will need them fixed either way.
Read the billing definition as well as the price. Ask each vendor what counts as a resolution, and whether a customer who leaves without replying is one.
Decide who owns it. Someone has to read the conversations every week, whichever route you choose.
Frequently asked questions
How much does an AI chatbot cost for a website in the UK?
On the three products we priced, between about 49p and £1.50 per resolved conversation on published prices, plus any plan or seat fees. Intercom Fin is $0.99 per outcome, Zendesk charges £1.30 per extra committed resolution (US$2.00 pay-as-you-go) on plans from £15 per agent a month, and Tidio Lyro starts at $32.50 a month for 50 conversations on annual billing.
How much does the AI model itself cost per conversation?
On our worked example of a three-question chat, between about 0.1p and 0.3p on small models, and about 2p to 3p on mid-range models such as Claude Sonnet 5.5 or OpenAI gpt-6.1-sol, using each provider’s published prices on 24 October 2026.
Is it cheaper to build your own AI chatbot?
The running costs are lower, often tens of pounds a month at 1,000 conversations. The build and the ongoing upkeep are not, because someone has to test the assistant, read its mistakes and move it to new models as old ones are retired. For most small businesses, the deciding factor is whether there is someone to own it.
What is a “resolution” in AI chatbot pricing?
Each vendor defines it. Intercom counts an outcome when a customer confirms their issue is resolved, does not ask for more help after the assistant responds, or the assistant completes a workflow including a handoff. Zendesk counts a request resolved by its AI agent without escalation to a human agent.
Do AI model prices change?
Yes. Google’s price list as captured on 23 October 2026 shows Gemini 3.8 Flash doubling in price on 1 January 2027, and Gemini 4 Argon doubling on 1 December 2026. Set your budget per conversation and recheck the price list regularly.
Flux Dynamics is a fractional CTO who builds. If you want to know what an assistant would cost on your own numbers, tell us what you run and we will work it out with you.