What does AI cost to run? Tokens and model pricing explained

AI models bill per token, and output costs more than input. Here is how the pricing works and how to estimate a monthly bill for a chatbot.

AI models are billed by the token, which is a chunk of text a little shorter than a word. You pay one rate for everything you send in (input) and a higher rate, five times as much on the current flagship models, for everything the model writes back (output). The model you pick moves the bill more than anything else: in the worked example below, the same website chatbot costs $3.90 a month on a small model and $312 on a top-tier one.

All three big providers publish their prices in US dollars per million tokens. The figures below were taken from the official pricing pages of Anthropic, OpenAI and Google on 18 November 2026 and checked again on 22 November 2026. They change often, so treat them as a snapshot.

  • A token is roughly four characters. Google’s rule of thumb is that 100 tokens is about 60 to 80 English words.
  • Output tokens cost five times as much as input tokens on the current flagship models from all three providers.
  • Top-tier models cost up to 80 times more per token than the cheapest ones, so match the model to the job.
  • Caching repeated text and batching non-urgent work can each cut the relevant part of the bill by half or more.
Claude API prices at a glance, November 2026
$0.10per million input tokens on Claude Haiku 5.5 (prompts up to 100,000 tokens)
$40per million output tokens on Claude Fable 5.2
50%discount for batch processing
30%more tokens for the same text on newer models' tokeniser

Source: Anthropic, Claude Platform pricing, checked 22 November 2026

What is a token?

Models don’t read words. They read tokens: short fragments of text that might be a whole word, part of a word or a punctuation mark. Google’s documentation says that for Gemini models a token is about four characters, and 100 tokens is about 60 to 80 English words. A one-page letter is therefore a few hundred tokens and a 20-page contract is in the low tens of thousands.

Token counts are not identical between models. Anthropic notes that its newer models use a tokeniser that produces approximately 30% more tokens for the same text than its earlier ones. So the same document can cost a different amount on two models even at the same headline price. Images and audio are converted to tokens too, by each provider’s own formula.

Why input and output are priced differently

Every request has two parts. Input is everything the model has to read: your instructions, any documents or product data you attach, and the conversation so far. Output is what it writes. Providers charge more for writing than for reading. On Claude Sonnet 5.5, for example, input is $2 and output is $10 per million tokens.

This has a practical consequence that surprises people. In a chat, the whole conversation is sent back to the model on every turn, so input tokens pile up quickly. In most business uses, input volume is much larger than output volume, even though each output token costs more.

How much do different AI models cost?

Each provider sells a range, from a large model for hard reasoning down to a small, fast one for simple, high-volume tasks. The gap is wide. Anthropic’s Claude Fable 5.2 and OpenAI’s gpt-6.2-astra both list at $8 input and $40 output per million tokens. The mid-range Claude Sonnet 5.5 and gpt-6.2-sol are both $2 and $10. At the bottom, Claude Haiku 5.5 and gpt-6-luna are $0.10 and $0.50 (for Haiku 5.5, on prompts up to 100,000 tokens). Google’s gemini-4-hydrogen is $1.50 and $7.50, and gemini-4-carbon is $2 and $10.

Output price per million tokens, selected models (USD)

List prices in US dollars before tax. OpenAI and Google figures are from their own pricing pages, linked under Sources. Google lists gemini-4-carbon at $20 from 1 December 2026.

Source: Anthropic, OpenAI and Google API pricing pages, standard rates on 18 November 2026

View data
 Output price per million tokens, selected models (USD)
Claude Fable 5.2$40
gpt-6.2-astra$40
Claude Opus 5.5$20
Claude Sonnet 5.5$10
gpt-6.2-sol$10
gemini-4-carbon$10
gemini-4-hydrogen$7.5
Claude Haiku 5.5$0.5
gpt-6-luna$0.5

That last note is worth a second look. Google’s pricing page lists gemini-4-carbon at $2 input and $10 output through 30 November 2026, rising to $4 and $20 from 1 December, with its context caching price doubling from $0.10 to $0.20 at the same time. Prices move in both directions, and a model you depend on can double in price with a few weeks’ notice. Build your budget with headroom.

Our usual advice: start with a mid-range model, measure quality on your real tasks, then try the cheaper tier. Small models now handle sorting, extraction and routine replies well. Keep the expensive ones for work where a wrong answer is costly.

How to estimate a monthly bill for a chatbot

Estimating a monthly AI bill

Step 1 of 5

Count volume

Estimate conversations or documents a month and turns per conversation.

Step 2 of 5

Size tokens

Measure input and output tokens for a few real examples.

Step 3 of 5

Multiply

Total input and output tokens, then apply each model's price per million.

Step 4 of 5

Add extras

Include tool calls, voice, hosting and staff review time.

Step 5 of 5

Cap it

Set a spending limit in the provider console and cap reply length.

Say you run a letting agency and put an AI assistant on your website. Assume 2,000 conversations a month, six exchanges each. On each exchange the model reads your instructions, some property information and the chat so far, averaging 2,000 input tokens, and writes a 250-token reply. These volumes are assumptions for illustration.

  • Input: 2,000 conversations × 6 turns × 2,000 tokens = 24 million tokens
  • Output: 2,000 conversations × 6 turns × 250 tokens = 3 million tokens
Model Input cost Output cost Monthly total
Claude Haiku 5.5 ($0.10 / $0.50) 24 × $0.10 = $2.40 3 × $0.50 = $1.50 $3.90
Claude Sonnet 5.5 ($2 / $10) 24 × $2 = $48 3 × $10 = $30 $78
Claude Fable 5.2 ($8 / $40) 24 × $8 = $192 3 × $40 = $120 $312
Monthly model cost for the same letting agency chatbot (USD)

Assumes 24 million input and 3 million output tokens a month, no caching or batching.

Source: Hahu worked example using Anthropic list prices, 22 November 2026

View data
 Monthly model cost for the same letting agency chatbot (USD)
Claude Haiku 5.5$3.9
Claude Sonnet 5.5$78
Claude Fable 5.2$312

Same traffic, an 80-fold difference in cost. For most customer service chat, the top model would be wasted money. If you’re weighing up whether a chatbot is worth having at all, we cover that in should your small business use an AI chatbot.

A document workflow

Now a back-office job. Say an accountancy practice extracts key fields from 5,000 supplier invoices a month. Each invoice plus instructions is 3,000 input tokens and the structured result is 300 output tokens. That is 15 million input tokens and 1.5 million output tokens. On gpt-6.2-sol at $2 and $10, the cost is 15 × $2 + 1.5 × $10 = $45 a month.

Caching and batching: the two discounts worth knowing

Caching applies when the start of every request is identical, such as a long set of instructions or a product catalogue. The provider stores that portion and charges far less to reuse it. Anthropic charges 1.25 times the input price to write a five-minute cache, then a fraction to read it: $0.10 per million on Claude Sonnet 5.5 against the normal $2. OpenAI lists cached input for gpt-6.2-sol at $0.10 against $2, and Google lists context caching for gemini-4-hydrogen at $0.075 against $1.50. In the chatbot example, most of the 24 million input tokens are repeated instructions and history, so caching could remove the bulk of the input cost.

Batching suits work that can wait. You submit jobs in bulk and collect results later. Anthropic and Google both state a 50% discount for their batch services, and OpenAI’s batch price table lists half its standard rates. The invoice example above would fall from $45 to $22.50. Anthropic says its batch and caching discounts can be combined.

What the token price leaves out

The model bill is rarely the biggest cost of an AI system. Budget separately for:

  • Building and maintaining it. Prompts, integrations, testing and monitoring take skilled time. See build or buy for how to weigh that against an off-the-shelf product.
  • Tools the model calls. Anthropic, for example, charges $10 per 1,000 web searches on top of tokens.
  • Voice. Phone agents add speech recognition, speech synthesis and telephony charges. We explain those in AI phone answering: what voice agents can and can’t do.
  • Hosting, logging and human review. Someone should read a sample of conversations every week.
  • Compliance work. If personal data is involved, run through our UK GDPR checklist for AI agents before launch.
  • Your own pricing. If you are building a product that resells AI features, usage costs shape your plans, which we cover in SaaS pricing models explained.

Three habits keep bills predictable: set a monthly spending limit in the provider’s console, cap the length of replies, and trim what you send. Sending a whole 50-page manual with every question when two paragraphs would do is the most common waste we see.

Common questions

Are the prices in pounds, and is VAT added?

API prices are listed in US dollars, so your cost in pounds moves with the exchange rate. VAT depends on which company invoices you. HMRC’s VAT Notice 741A says that when a UK VAT-registered business buys services from a supplier outside the UK, the supply is generally treated as made in the UK and the customer accounts for the VAT under the reverse charge, which has no net cost if you can recover all your input tax. Check your invoices and ask your accountant, as this is general information and not tax advice.

Is a ChatGPT or Claude subscription the same thing?

No. A monthly subscription gives a person access to a chat app at a flat fee. Token pricing applies when software you build calls the model through its API, which is how chatbots, agents and automated workflows run.

The short version

Pick the cheapest model that does the job well on your real examples, cache what repeats, batch what can wait and set a spending cap. For context on how other firms are using these tools, see how UK small businesses are using AI. If you want a costed estimate for your own product, our SaaS product development team can work one through with you.

Sources

→Keep reading

Related guides.

All insights
Software reviews

Cursor vs GitHub Copilot vs Claude Code: which AI coding tool fits your team?

Cursor, GitHub Copilot and Claude Code compared on published prices, plans, usage limits and developer survey data, with a best fit verdict for UK… Read
AI explained

RAG explained: how “chat with your documents” AI really works

Retrieval-augmented generation lets AI answer from your own files. How RAG works, what it costs, why it still gets things wrong and when to… Read
AI explained

Open-source AI models explained: Llama, Gemma, DeepSeek and the rest

What open-weight AI models are, how the Llama, Gemma, DeepSeek, Qwen, Mistral and gpt-oss licences differ, and when a UK business should use one. Read

→Get in touch

Got a question? Ask us.

Tell us what you’re trying to do. Every project starts with a free audit: someone who does the work replies with a few ideas, an honest view and a written plan with prices.

What happens next

  1. 1
    You send the formIt takes about two minutes.
  2. 2
    We do our homeworkWe look at your website and your market.
  3. 3
    You get a short planWhat we’d do first, with prices. Yours to keep.
  • Free, with no obligation
  • We won’t add you to a mailing list

We only use your details to reply to you. See our privacy policy.