What does it cost to add AI features to business software through an API?
There are two bills. The build starts at £799 with us, scoped per project. The AI provider then charges per million tokens, from $0.10 in and $0.50 out on small models to $4 in and $20 out on a larger one. Drafting 1,000 customer replies costs about $0.30 on a small model and about $6 on a mid-range one.
| Task, with the size we assumed | Small model ($0.10 in, $0.50 out) | Mid-range model ($2 in, $10 out) | Larger model ($4 in, $20 out) |
|---|---|---|---|
| Draft a reply to an enquiry (1,500 tokens in, 300 out) | $0.30 per 1,000 replies | $6 | $12 |
| Pull the details out of a form or email (2,000 in, 200 out) | $0.30 per 1,000 items | $6 | $12 |
| Summarise a 10-page document (7,000 in, 500 out) | $0.95 per 1,000 documents | $19 | $38 |
| An 8-message chatbot conversation (22,000 in, 1,200 out) | $2.80 per 1,000 conversations | $56 | $112 |
Our arithmetic from the providers’ published prices per million tokens, checked 17 November 2026. The token counts are assumptions, stated so you can redo the sum. Prices are in US dollars because that is how the providers charge.
When business software “uses AI”, it is almost always sending text to a model run by OpenAI (the company behind ChatGPT), Anthropic or Google, and getting text back. That link is an API, and each call is metered. This guide shows how the meter works and what it costs today. If what you want is a chatbot on your website, our UK chatbot cost guide prices that on its own, and our guide to AI automation for small businesses covers which jobs are worth automating in the first place.
If you would rather hand this to us, we offer our custom business software service from £1,499.
How AI APIs charge: tokens in, tokens out
Providers do not charge per question. They charge per token, a chunk of text a little shorter than a word. Google’s rule of thumb for its Gemini models is that a token is about 4 characters and 100 tokens is about 60 to 80 English words.
Three things follow from that:
- Input and output are priced separately, and output costs more. On every model in the table below, a token written by the model costs at least five times a token you send.
- Everything you send counts. The instructions, your price list, the customer’s message and the earlier messages in a chat are all input, every time.
- The same text is not the same number of tokens everywhere. Anthropic notes that its newer models use a tokenizer that produces roughly 30% more tokens for the same text, so compare models on the cost of your task, not only on the headline price.
Published prices per million tokens
These are standard prices from each provider’s own pricing page, for three models from each, from small to larger:
| Provider and model | Input, per 1M tokens | Output, per 1M tokens | Note |
|---|---|---|---|
| OpenAI gpt-6-luna | $0.10 | $0.50 | |
| OpenAI gpt-6.1-terra | $1.00 | $5.00 | |
| OpenAI gpt-6.1-sol | $2.00 | $10.00 | |
| Anthropic Claude Haiku 5.5 | $0.10 | $0.50 | For prompts up to 100,000 tokens |
| Anthropic Claude Sonnet 5.5 | $2 | $10 | |
| Anthropic Claude Opus 5.5 | $4 | $20 | |
| Google gemini-3.5-flash-lite | $0.30 | $2.50 | |
| Google gemini-3.8-flash | $0.75 | $3.75 | Listed to rise to $1.50 and $7.50 on 1 January 2027 |
| Google gemini-4-carbon | $2.00 | $10.00 | Listed to rise to $4.00 and $20.00 on 1 December 2026 |
Look at the last column: two of the nine rows already carry a dated price rise, so re-check this table after 1 January 2027. Prices and model line-ups change, so a well-built feature keeps the model as a setting that can be switched. Then a price rise means a small change, not a rebuild.
Worked example: an enquiry assistant for a heating firm
Take a heating firm that receives 600 web and email enquiries a month. It wants the postcode, boiler type and urgency pulled out of each one, and a reply drafted for the office to approve. That is two calls per enquiry, using the sizes in the first table.
- On a small model: $0.0003 + $0.0003 = $0.0006 per enquiry, so 600 enquiries cost about $0.36 a month.
- On a mid-range model: $0.006 + $0.006 = $0.012 per enquiry, so about $7.20 a month.
- On a larger model: $0.024 per enquiry, so about $14.40 a month.
Even at ten times the volume, the provider’s bill stays under $150 a month on the larger model. For most small firms the API fee does not decide the project. The build does.

What makes the running cost climb
- Long conversations. A chat resends its history with every message. In our 8-message example, a 2,000-token brief goes out eight times, which accounts for 16,000 of the conversation’s 22,000 input tokens.
- Big documents. Sending a whole manual when the answer is on one page pays for every page.
- Using a large model for simple jobs. The first table shows a 40-fold gap between the small and larger tiers for the same task.
- Add-on tools. OpenAI lists web search at $10 per 1,000 calls plus the tokens it reads, and file search at $2.50 per 1,000 calls.
- Regional processing. OpenAI charges a 10% uplift on its data residency endpoints for models released on or after 5 March 2026.
Four ways to pay less for the same result
- Match the model to the job. Start on the smallest model, test it on 50 real examples, and move up only where it fails.
- Cache the part that repeats. Instructions and price lists are the same on every call. OpenAI lists cached input on gpt-6-luna at $0.01 per million tokens against $0.10 uncached. Anthropic prices cache hits on Claude Sonnet 5.5 at $0.10 per million against $2, though writing to the cache costs a little more than normal input.
- Batch what can wait. Anthropic’s Batch API gives a 50 percent discount on input and output for work processed in the background. Google lists its Batch API at a 50% cost reduction, and OpenAI’s batch prices are half its standard prices on the models we checked.
- Send less and ask for less. Trim the history, send the relevant page and not the whole file, and cap the length of the answer. Output is the expensive side.
The build fee: what you pay us and what changes it
The provider’s meter is one bill. The other is the work of building the feature into your software. Our AI-Powered Tools & Automation service starts at £799 and covers custom chatbots, intelligent automation, document processing and smart alerts, scoped per project. These are the things that move the price:
- How many systems it touches. Reading from your inbox and writing to your job system are two connections. Our API Development & Integration service starts at £499, and our API integration cost guide explains how those are priced.
- Whether a person approves the result. A draft for the office to check is simpler than a reply sent straight to a customer.
- The state of your data. A tidy price list is quick to use. Years of mixed spreadsheets need cleaning up first.
- Testing. A feature customers will see needs testing against awkward inputs, not only the easy ones.
Every price we charge is on our pricing page.
Costs by situation: which model tier to budget for
| Situation | Tier to start on | Why |
|---|---|---|
| Sorting and tagging incoming enquiries | Small | Short input, a one-word answer, easy to check |
| Drafting replies a person approves | Small, then mid-range if the tone is off | The person catches mistakes, so cost can lead |
| Reading long tenders, contracts or reports | Mid-range or larger | Mistakes are costly and volumes are low |
| A customer-facing chatbot | Mid-range, with caching | Nobody checks each answer, and the same brief is resent on every message |
| Bulk overnight jobs, such as rewriting 5,000 product descriptions | Small, through a batch API | Nothing is waiting on the answer, so take the 50% discount |
If you are still choosing between an off-the-shelf AI tool and a built one, our AI tool stack guide for small businesses lists what to install before you build anything.
Data protection: check this before you send customer data
What happens to the text you send differs by provider and by plan:
- OpenAI says data sent to its API is not used to train or improve its models unless you explicitly opt in, and that abuse monitoring logs are kept for up to 30 days by default.
- Google lists its free tier as used to improve its products and its paid tier as not. A free tier is for testing with made-up data, not for customer records.
- The ICO publishes guidance on AI and data protection for organisations, including what to consider in a data protection impact assessment. It notes the guidance is under review following the Data (Use and Access) Act.
You stay responsible for your customers’ data whichever provider processes it. Our GDPR-compliant software guide covers the questions to put to any supplier.
Questions to ask before you pay for an AI feature
- Which model, and why that one? Ask what the task costs on a smaller model.
- What is the fee per task, worked out from real examples of my data?
- Whose provider account is it? If it is yours, you see the bill and keep control if you change developer.
- Can the model be swapped without a rebuild when prices or models change?
- What happens when the AI is wrong, and who sees it?
- Is usage logged per feature, so a jump in the bill can be traced?
The short version
Budget for the build first: from £799 with us, moving with the number of systems and how much checking the result needs. Then work out the fee per task from the published price per million tokens. For most small-business uses it is a small monthly sum, provided the model fits the job and the repeated part is cached.
Sources
- OpenAI, API pricing (model prices, cached input, batch prices, tool prices, regional processing uplift): developers.openai.com/api/docs/pricing
- Anthropic, Claude pricing (model prices, cache pricing, Batch API discount, tokenizer note): platform.claude.com/docs/en/about-claude/pricing
- Google, Gemini API pricing (model prices, scheduled changes, free and paid tier data use, batch): aistudio.google.com/docs/pricing
- Google, Gemini API, Understand and count tokens (token rule of thumb): aistudio.google.com/docs/tokens
- OpenAI, Your data (training and retention of API data): developers.openai.com/api/docs/guides/your-data
- ICO, Guidance on AI and data protection: ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection
All provider prices and terms were checked on 17 November 2026 and change often; confirm them on the pages above before you budget. Dollar prices are shown as published and are not converted. The cost-per-task figures are our own sums from those prices and the stated token counts. This guide is not legal advice.
Get the UK Website Project Brief Template (free)
A two-page template for briefing any UK web designer: sitemap, integrations, audience and deliverables. Cuts revision rounds in half.
No spam. UK GDPR compliant. We only use your email to send this resource, unless you tick the box to hear from us.
Frequently asked questions
How much does it cost to use the ChatGPT API in business software? +
OpenAI charges for its API per million tokens, not per month. On its pricing page on 17 November 2026, gpt-6-luna was $0.10 per million input tokens and $0.50 per million output tokens, and gpt-6.1-sol was $2 and $10. Drafting 1,000 short customer replies works out at about $0.30 to $6 on those two models.
What is a token in AI pricing? +
A token is the unit AI providers bill in: a chunk of text a little shorter than a word. Google's rule of thumb for its Gemini models is that a token is about 4 characters and 100 tokens is about 60 to 80 English words. You pay for the tokens you send and, at a higher price, the tokens the model writes.
How much does one AI chatbot conversation cost in API fees? +
On our assumptions, an 8-message conversation that resends a 2,000-token brief each time uses about 22,000 input tokens and 1,200 output tokens. That is about $0.003 on a small model priced at $0.10 in and $0.50 out, and about $0.06 on a mid-range model priced at $2 in and $10 out.
Is the API bill or the build the bigger cost? +
For most small businesses, the build. Our AI-Powered Tools & Automation service starts at £799, scoped per project. In our worked example, a firm handling 600 enquiries a month would pay the AI provider about $0.36 to $14.40 a month, depending on the model.
How can I cut AI API costs without losing quality? +
Use the smallest model that passes your tests, cache the instructions that repeat on every call, send only the text the task needs, cap the length of the answer, and run jobs that can wait through a batch API. Anthropic and Google both list a 50% discount for batch processing.
Is it safe to send customer data to an AI API? +
It depends on the provider and the plan, so check before you send anything. OpenAI says API data is not used to train its models unless you opt in. Google lists its free tier as used to improve its products and its paid tier as not. You remain responsible for the data under UK data protection law.


