Contents7
The cost of AI in a business does not work like the cost of software. As soon as an AI is connected to your tools through an API (the access point a program calls, without going through the chat interface), there is no per-user subscription any more: each call is paid for according to the amount of text processed. This guide, from our Business AI: private and at the right cost section, explains what you pay for and how to forecast it. Our guide to integrating AI in business places this question of price between choosing the model and setting the rules.
In short
An AI used through an API is paid for by the token, on input and on output; on the Anthropic price list checked on 4 October 2026, output costs five times as much as input.
The context sent back with every call, the declared tools and an agent’s turns often weigh more than the answer itself.
Discounts stack: at Anthropic, a cache read costs 10% of the input price on most models, and batch processing takes off 50%.
A bill is forecast use by use, with tokens measured on real examples, and then kept in check with caps.
How an AI bill is calculated when you use an API
An AI used through an API is paid for by the amount of text processed, counted in tokens.
Definition
The token
A piece of a word. Anthropic’s pricing documentation gives an order of magnitude: in English, a token is about 4 characters, or three quarters of a word, and the exact count varies with the language and the type of content. French text, for instance, is not split up the way English text is.
Two meters run on every call: input tokens (everything you send: the instruction, the document, the conversation history) and output tokens (what the model writes). Output costs more. Here is Anthropic’s public price list for its current models, checked on 4 October 2026, in dollars per million tokens:
| Model | Input | Output |
|---|---|---|
| Claude Haiku 4.5 | $1 | $5 |
| Claude Sonnet 5.5 | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
| Claude Fable 5.1 | $10 | $50 |
Across this whole range, output costs five times as much as input. These prices move, and each vendor has its own price list: check the current ones when you do your calculations.
What inflates a bill without anyone noticing
The context sent back with every call
A language model has no memory between two calls. For an assistant to “remember” a conversation, the application sends it the whole history with every message. By the tenth exchange of a discussion, you have paid for the first message ten times. The same goes for a fifty-page document pasted into every request: you pay to have it read at every question. There are two ways round this: the cache, described below, and RAG, which sends the model only the useful passages of your documents.
Tools and system prompts
An assistant that can call tools (search a CRM, read a file, run a web search) sends their description with every request. According to Anthropic’s price list, the system prompt added as soon as a tool is declared weighs a few hundred tokens per request (286 for Claude Opus 5.5 and Claude Sonnet 5.5, 496 for Claude Haiku 4.5), before the description of your own tools; a web search costs an extra $10 per 1,000. Small amounts, but visible across tens of thousands of calls.
An agent’s turns
An agent is a program that chains several calls to the model to reach a goal: it thinks, calls a tool, reads the result, and starts again. Each turn sends back the accumulated context. A task that takes eight turns does not cost eight times a simple answer. It often costs more, because the context grows with every turn. This is a classic hidden cost of agent projects, and it can be capped: a maximum number of turns, a budget per task, a clean stop when the cap is reached.
Moving to a new model generation
Even at the same unit price, a new generation can count tokens differently.
Watch out
A model migration has to be recalculated, not assumed
Anthropic states that its models from Claude 4.7 onwards use a different way of cutting text into tokens, which produces about 30% more tokens for the same text, the exact gap depending on the content. Its documentation on token counting recommends re-counting your requests with the target model rather than reusing measurements made on the old one.
How to bring the bill down: three levers
1. The right model for each task (routing)
Routing means sending each task to the cheapest model that gets it right. Sorting an incoming email or extracting an invoice number does not need the most powerful model in the catalogue. Writing a sales page or settling an ambiguous case, perhaps. On the price list above, the input price goes from 1 to 4 between Claude Haiku 4.5 and Claude Opus 5.5, and from 1 to 10 up to Claude Fable 5.1. Anthropic, incidentally, recommends Haiku for simple tasks, Sonnet for most production uses and Opus for the most complex reasoning.
The choice of vendor matters as much as the choice of model.
From the field
On the translation pipeline we operate for Palmora Property, a six-language real estate portal we work with, switching vendor, from DeepL to Gemini, cut the cost by a factor of 60. That ratio applies to that pipeline: it shows how far a trade-off can go, not what you will save.
A routing table fits in a few lines: the task, the default model, the fallback model, the reason for the choice. Kept in a configuration file, it can be revised without touching the automations. You still have to prove that a cheaper model is enough, which is what our guide to evaluating an AI model before putting it into production is about.
2. Do not pay twice for the same work (caching)
Vendors offer a prompt cache: the fixed part of a request (instructions, reference document) is stored and read back at a reduced price. On Anthropic’s price list, a cache read costs 10% of the normal input price on most models, and less still on Claude Opus 5.5 (5%) and Claude Fable 5.1 (2.5%). Writing to the cache costs a quarter more than normal input for a five-minute duration: according to the vendor, the cache pays for itself from the first re-read within that time.
Caching on the application side matters just as much: never ask the AI again for something it has already produced. On the same Palmora Property translation pipeline, each text segment gets a fingerprint, and only new or modified segments go back for translation. After the first pass, 95% of segments come out of the cache without calling the AI again.
3. Agree to wait (batch processing)
Anything that does not need an immediate answer can go through batch processing: you send a bundle of requests and collect the results later. At Anthropic, the batch API applies a 50% discount on input and output, which can be combined with the cache. Classifying a ticket history or enriching a product catalogue can wait. One caveat for sensitive data: Anthropic’s page on data retention states that batches are kept for 29 days and are not covered by the zero-retention agreement.
How to forecast the bill before you receive it
The forecast is done use by use, in five steps:
- List the uses, each on its own line: “sorting incoming emails”, “summarising meeting minutes”.
- Estimate the volume of requests per month from your real figures (emails received, tickets opened).
- Measure the tokens on real examples, on input (instruction, content sent, history) and on output (expected length of the answer). Anthropic offers an endpoint that counts the tokens of a message before it is sent; this count is not billed, and it is an estimate that can differ slightly from the real count.
- Apply the price of the chosen model from your routing table, then the discounts that apply: caching, batches, or none.
- Add the human time for review and correction, which tokens do not show and which can weigh more than they do.
An example, with assumptions of our own and the prices checked on 4 October 2026: a team hands the sorting of its incoming emails to Claude Haiku 4.5.
A team’s email sorting, from volume to bill
- The volume
2,000 emails a month.
- Tokens per request
About 1,500 sent (the instruction and the message), 50 received (a category).
- The month’s total
3 million input tokens, 100,000 output tokens.
- The bill
$3 for input, $0.50 for output.
The same sorting given to Claude Opus 5.5 would cost more than four times as much (a unit price four times higher, and a different way of cutting text that counts more tokens for the same text), with no guarantee of better classification. Anthropic publishes a similar calculation for customer support on its price list: about $37 for 10,000 tickets with Haiku 4.5, on an average of 3,700 tokens per conversation.
An estimate like this gets corrected after the first week, with real measurements.
How to keep the bill under control
An AI bill drifts if nobody looks at it. Three habits:
- One API key per use, so you know what each automation costs, and not just the month’s total.
- Caps and alerts at the vendor or in the gateway that centralises the calls: an agent stuck in a loop must not be able to use up a month’s budget in one night.
- A quarterly review of the routing table: prices change, new small models arrive, and a task that needed a large model six months ago may not need one any more.
A gateway that gives access to several models helps you keep these habits: it centralises keys and caps, and lets you change models without rewriting the integrations. Check what it charges on top of the models, and remember that it sees all your requests go through: it is one more intermediary to write into your contracts and your record of processing activities. For a concrete example, see OpenRouter for a marketing agency: one key and one cap per client give the real AI cost of each account.
Where to start
Start from the use that consumes the most, measure it for a week, then test a cheaper model on your own examples. The same sorting of tasks lightens the environmental footprint too, since a smaller model uses less energy: that is the subject of our article on the energy and water consumed by AI. If your requests contain personal data, also look at where they go depending on the vendor and the level of control you choose. This guide does not cover per-user subscriptions (ChatGPT, Copilot, AI options in business software), which are compared differently.
To get an audit of your current bill, with a prioritised savings plan and a benchmark on your own content, see our AI at the right cost offer.
Frequently asked questions
What is a token, and why watch the length of answers?
A token is a piece of a word: about four characters in English according to Anthropic, and the exact count varies with the language and the content. On that vendor’s price list checked on 4 October 2026, an output token costs five times as much as an input token. So ask the model for short, structured answers, a category or a field, rather than paragraphs.
How do you estimate the cost of an AI agent before putting it into production?
Count two bills: the build (scoping, connection to your tools, testing) and the running (tokens, hosting, monitoring, human review). The second varies more than a simple call, because it depends on the number of turns per task and on a context that grows with each turn. Run the agent on a few dozen real tasks, measure the tokens consumed, multiply by the expected volume, then set a cap per task before opening it to the whole team.
What is model routing, and how much does it save?
Routing sends each task to the cheapest model that gets it right, according to a table written in advance. The saving has no universal percentage: it depends on your tasks and on the price gap between the models involved, for example from 1 to 4 on input between Claude Haiku 4.5 and Claude Opus 5.5 on Anthropic’s price list. Only tests on your own examples tell you whether the cheaper model is enough.
Sources
- Anthropic’s pricing documentation (new tab)platform.claude.com
- Its documentation on token counting (new tab)platform.claude.com
- Anthropic’s page on data retention (new tab)platform.claude.com
Jim · founder of Brain Plus AI
Also in business ai
Practical guide9 min
Evaluating an AI model before putting it into production
Public leaderboards do not tell you whether a model will do your work well. This guide gives a method for testing it on your own tasks: quantified criteria, real examples, scoring, cost, and a new test whenever the model changes.
Choose and price9 min
Private AI in business: the four levels of control
A "private" AI is not a product, it is a degree of control over your data. We distinguish four levels, from an API used as is to a model installed on your premises, and a simple way to choose task by task.
Understand9 min
Enterprise RAG: an AI that answers from your documents
RAG connects an AI model to a company's documents so that it answers by citing its sources. How it works, when it is useful, why it sometimes misses the point, and what to set up before opening it to the team.