How much does an AI chatbot cost to run? Claude prices in pounds

Most people asking what an AI chatbot costs are braced for a big number. The model bill is usually the smallest line in the budget. A support assistant handling 10,000 conversations a month can cost less than £15 in model fees. The same assistant on the most capable model could cost over £2,000.

That gap is the whole story. Which model you pick, and how you feed it, moves the bill by more than a hundred times. The build, the integrations and the people who keep it honest cost far more than either figure.

This guide does the token maths in the open, using Anthropic’s published Claude prices. Every assumption is stated, so you can swap in your own numbers.

How is an AI chatbot priced?

You pay per token, not per conversation or per seat. A token is a small chunk of text, usually part of a word. On Anthropic’s current tokeniser, one million tokens is roughly 555,000 words, according to its models overview.

Two kinds of token are billed separately:

• Input tokens are everything you send the model: your instructions, the help content you retrieved, the conversation so far and the customer’s new message.
• Output tokens are everything the model writes back. That includes any reasoning it does before answering, which is billed as output even when you don’t show it, per the thinking documentation.

Output costs five times as much as input on every current Claude model. But a chatbot sends far more than it receives. In our example below, input makes up about 80% of the bill.

What do the Claude models cost?

These are Anthropic’s list prices in US dollars per million tokens. We checked the Claude pricing page and the API pricing documentation on 20 October 2026.

Anthropic describes Haiku 5.5 as built for high-volume, latency-sensitive tasks such as classification and routing. Sonnet 5.5 is pitched as the best mix of speed and intelligence. Opus 5.5 is its recommended starting point for most workloads. Fable 5.1 is for demanding reasoning and long-running agent work.

A few details change the sums:

• Haiku 5.5 has a long-prompt tier. A request whose prompt runs over 100,000 tokens pays $0.50 input and $2.50 output. A support chatbot should stay well under that.
• Prompt caching lets you store a repeated chunk of prompt, such as your instructions. Writing it costs 1.25 times the input price for a five-minute cache. Reading it back costs a fraction of the input price, as the table shows.
• Batch processing halves the price of input and output. Answers come back later rather than instantly, so it suits overnight jobs, not live chat.
• US-only processing costs 1.1 times the standard price, and Opus 5.5’s fast mode costs twice the standard price.
• Web search, if your bot uses it, costs $10 per 1,000 searches on top of tokens.

What assumptions sit behind the worked examples?

Picture a support assistant on a UK subscription business’s website. It answers questions from help articles, and hands anything awkward to a person. Each conversation looks like this:

Each turn resends the instructions, fresh help content and the conversation so far. Old help passages are dropped once a turn is answered. That gives 22,800 input tokens and 1,200 output tokens per conversation.

For pounds, we used the European Central Bank’s reference rates for 19 October 2026. Its euro rates for the dollar and the pound give $1 = £0.75. Anthropic bills in dollars, so your card provider’s rate and any fees will move the pound figures slightly.

What does 1,000, 10,000 or 100,000 conversations a month cost?

With no optimisation

With the instructions cached

Now cache the 3,000-token instruction block, so every turn reads it from the cache. Nothing else changes.

Caching takes roughly 40% off every model here. It ignores the first write of each cache, which is small at steady traffic.

One caveat at low volume. The default cache lives for five minutes and refreshes each time it’s read. At 1,000 conversations a month, roughly 33 a day, it will often go cold between chats. A one-hour cache costs twice the input price to write. At that volume the whole bill is small, so don’t lose sleep over it.

How do you bring the bill down?

Route easy questions to a smaller model

Most support traffic is repetitive: delivery times, opening hours, password resets. Haiku 5.5 costs a twentieth of Sonnet 5.5 for both input and output.

Suppose 80% of conversations go to Haiku 5.5 and 20% go to Sonnet 5.5, both with caching. At 10,000 conversations a month, that’s $84 (£63). Sending everything to Sonnet 5.5 costs $348 (£261).

The router can be Haiku itself, classifying each new message before anything else happens. Classification and routing are exactly the jobs Anthropic says Haiku is built for. Test the split on your own past conversations before you trust it.

Cache everything that repeats

Put stable content first in the prompt: instructions, policies, tool definitions and any help content every conversation needs. Caching only works on a stable prefix of at least 512 tokens on these models, per the prompt caching documentation. Anything that changes per customer goes after it.

Keep retrieved context tight

Retrieval is the biggest uncached cost in our example. Three tight passages beat ten loose ones, for cost and for accuracy. Cut retrieved help content from 2,000 to 800 tokens a turn. At 10,000 conversations, Sonnet 5.5 with caching drops from $348 to $252 a month.

Tighter context also means fewer chances for the model to quote the wrong article. Getting there depends on clean, well-split help content. Our guide to getting your data ready for AI covers that.

Cap the reply and the reasoning

Output is the expensive token. Set a maximum reply length, ask for short answers and use a lower effort setting for simple questions. Reasoning tokens are billed as output, so a model that thinks hard about opening hours is wasting money.

Batch what isn’t live

Summarising yesterday’s chats, tagging them by topic or drafting help articles from common questions can all wait. Run them through the Batch API at half price.

When does a bigger model pay for itself?

Model spend is rarely the deciding cost. A conversation that ends in a wrong answer, an angry customer or a human handover costs far more than the tokens.

Here’s a simple way to decide. At 10,000 conversations a month, Opus 5.5 costs £261 more than Sonnet 5.5 in our cached example. Say a human handover costs you £5 of staff time. Opus 5.5 pays for itself if it prevents about 52 extra handovers a month. That’s around one conversation in every 200.

That £5 is a placeholder, so use your own figure. The point is to compare models on resolved conversations, not on price per token. Run both on a few hundred of your real past questions, mark the answers and let the numbers decide.

Go bigger when questions need judgement across several sources, when mistakes are costly or when the bot takes actions in other systems. Stay small when the answer sits in one help article.

What else does a chatbot cost?

This is where the money really goes. Even on Opus 5.5, our 10,000-conversation month costs £522 to £864 in model fees. The surrounding work is the larger bill.

• Building it. Conversation design, retrieval over your help content, the chat interface and the hand-off to a person. A demo takes days; something you’d put in front of customers takes much longer.
• Integration. Order status, account details and booking slots live in other systems. Connecting them safely, read-only where possible, is usually the slowest part.
• Evaluation. You need a test set of real questions with agreed good answers. Run it before launch and after every prompt, model or content change.
• Monitoring. Someone has to read a sample of conversations each week. Watch for wrong answers, failed handovers and questions nobody wrote an article for.
• Human handover. A bot that can’t pass a customer to a person, with the conversation attached, will damage trust fast. That means staff, rotas and a helpdesk that can receive it.
• Content upkeep. The bot is only as current as your help articles. Someone has to own them, by name.
• Data protection. Chats contain personal data, so you need a lawful basis, a processing agreement and a retention rule. Our guide to AI and GDPR in the UK covers the detail.
• Maintenance. Models are retired on a published schedule, and prices change. Anthropic commits to keeping Haiku 5.5 available until at least 7 October 2027. Budget for a tested migration every year or so.

If your chatbot is customer-facing, read AI customer service for small business first. It argues that most small businesses should put AI behind the inbox before putting it in front of customers.

A quick way to estimate your own bill

1. Count your monthly conversations from your helpdesk or chat logs.
2. Pull 20 real conversations and count their turns and length.
3. Estimate tokens per conversation using the assumptions table above, adjusted to your numbers.
4. Multiply by the price list, then convert at today’s rate.
5. Add 50% headroom for retries, longer chats and growth.

If the answer is a few pounds a month, the model choice is about quality, not cost. If it runs to thousands, routing and caching deserve real engineering time.

How we build and run these

We build AI products end to end at Contic, with senior engineers reviewing and owning everything that ships. We use AI in our own work, responsibly. Our founder was a founding engineer at AI Music, which Apple acquired. We’re also a Databricks and AWS partner.

The closest thing we’ve built to a support bot is for Everyone’s Invited. It’s a charity that was receiving over 100 personal stories a day. Our system automatically removes names, dates and locations before staff and volunteers publish each one. It has handled over 50,000 submissions and saved around 12,500 hours of review.

After launch, TotalCare runs the product for a fixed monthly price. That covers hosting, third-party services and support, so you can get on with the business. You own the code, and the accounts are yours at launch.

If you’re weighing up a chatbot, talk to us about AI. The first chat is free, and so is an architectural review. We’ll tell you which model, if any, your volumes justify. You can also score your readiness first with our AI readiness scorecard.