Claude API pricing explained for small businesses

The four things that set the price

1. 01The model. Anthropic offers a large, most capable model, a balanced middle model and a small, fast model. The small model costs a fraction of the large one per token.
2. 02Input tokens: everything you send, including instructions, documents and chat history.
3. 03Output tokens: everything Claude writes back. These cost several times more than input tokens.
4. 04Discounts: prompt caching makes repeated input much cheaper, and the batch service processes non-urgent jobs at a lower rate. At the time of writing, Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, Claude Sonnet 5.5 costs $2 and $10, Claude Opus 5.5 costs $4 and $20, and Claude Fable 5.1 costs $10 and $50. A cache read costs 2.5 to 10 per cent of the normal input price depending on the model, and the Batch API halves both input and output prices.

A worked example

Say you automate replies to 2,000 customer emails a month. Each request sends 1,500 tokens of instructions and history plus a 300-token email, and gets a 250-token reply back. That's about 3.6 million input tokens and 0.5 million output tokens a month.

Multiply each by the model's rate per million tokens and add them together. Then work out the same job on the smaller model, and with the instructions cached. The difference between the most expensive and cheapest set-up is often several times over. At Anthropic's list prices on 3 November 2026, that's about $12.20 a month on Claude Sonnet 5.5 ($2 per million input tokens, $10 per million output tokens), about $61 on Claude Fable 5.1 and about $0.61 on Claude Haiku 5.5. If 1,500 tokens of each request were read from the cache, Sonnet 5.5 drops to around $6.50 before cache write charges. Anthropic bills in US dollars, so the cost in pounds depends on the exchange rate.

API or a Claude subscription?

A Claude Team or Enterprise subscription is priced per person and suits staff using Claude directly for writing, research and analysis. The API is for software and automations that call Claude in the background.

Many businesses use both: subscriptions for the team, and the API for the automations. Keep them separate in your budgeting, because they're billed differently.

Keeping the bill under control

Set spend limits in the Claude Console, track usage by API key so you can see which automation costs what, and review the busiest jobs each month. Shorter prompts, caching and choosing the smallest model that does the job well are usually the biggest savings.

How Claude API pricing works

The Claude API is Anthropic's pay-as-you-go service for software that calls Claude. There's no monthly fee for the API itself. You add credit or set up billing in the Claude Console, and you pay for the tokens each request uses.

A token is a small chunk of text, roughly three quarters of an English word on average. Every request has input tokens, which is everything you send, and output tokens, which is everything Claude sends back. Anthropic publishes a rate per million tokens for input and for output, and each Claude model has its own rates. At the time of writing (November 2026), Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, Claude Sonnet 5.5 costs $2 and $10, Claude Opus 5.5 costs $4 and $20, and Claude Fable 5.1 costs $10 and $50.

So the pricing structure is simple at heart: tokens in multiplied by the input rate, plus tokens out multiplied by the output rate. Everything else, such as caching, batch processing and tool use, adjusts one side of that sum.

What counts as input tokens

Input tokens include the system prompt, any tool definitions, the chat history you send back each turn, documents and images you attach, and the user's message. In many business automations, the repeated system prompt and context make up most of the input, not the customer's actual question.

What counts as output tokens

Output tokens are the words Claude writes back, plus any thinking tokens if you turn on extended or adaptive thinking. Thinking is billed as output, so a request that reasons at length costs more even if the final answer is short. Output is priced higher than input on every model, so asking for shorter answers is one of the easiest savings.

Claude models and their pricing tiers

Anthropic groups its Claude models into tiers. The names and version numbers change as new models are released, so always check Anthropic's pricing page for the current line-up. At the time of writing (November 2026), the current models are Claude Haiku 5.5, Claude Sonnet 5.5 and Claude Opus 5.5, plus Claude Fable 5.1, a newer tier above Opus for the most demanding reasoning and long agentic work.

• Older models often stay available for a while after a new one is released, sometimes at different rates.
• A newer model can use a different tokenizer, so the same text may count as a different number of tokens. Test before you switch.
• Pick the model per job, not per business. One automation can use Haiku while another uses Opus.

Claude Opus

Claude Opus is the most capable of the three core models and costs more per token than Sonnet or Haiku. Anthropic's Fable models sit above it for the hardest work, at a higher price. It suits complex reasoning, difficult analysis and long, multi-step work where quality matters more than cost. Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens.

Claude Sonnet

Claude Sonnet is the balanced middle tier. For many business tasks, such as drafting replies, summarising documents and extracting data, it's the recommended starting point because it offers strong quality at a lower price than Opus. Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens.

Claude Haiku

Claude Haiku is the small, fast model. It's the cheapest per token and works well for sorting, tagging, routing and short replies at high volume. Try Haiku first on simple jobs and only move up if the quality isn't there. Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, rising to $0.50 and $2.50 for longer prompts.

Discounts and extras that change your Claude API costs

The base rate per million tokens is only part of the picture. Several features change the standard token costs, some downwards and some upwards.

Prompt caching

Prompt caching stores repeated context, such as a long system prompt or a policy document, so later requests can reuse it. Writing to the cache costs a little more than standard input, but cache reads are charged at a much lower rate than the standard input price. If the same text goes out with hundreds of requests, caching is often the biggest single saving. At the time of writing, a cache write costs 1.25 times the base input price for the standard 5-minute cache, or 2 times for a 1-hour cache. Cache reads cost 0.1 times the base input price on most models and 0.05 times on Claude Opus 5.5 and Sonnet 5.5, so a Sonnet 5.5 cache read costs $0.10 per million tokens against $2 for standard input. The cache timer resets each time it's read.

The Claude Batch API

The Batch API processes requests asynchronously. You send a batch of requests, and Anthropic returns the results later rather than straight away, in exchange for a lower rate on both input and output tokens. It's ideal for work that can wait, such as overnight document processing, bulk tagging, report drafting or evaluating prompts. It isn't suitable for live chat or anything a customer is waiting on. Anthropic's Batch API charges 50% of the standard price for both input and output tokens. Most batches finish within an hour, and any request not processed within 24 hours expires and isn't charged. Batch processing discounts can usually be combined with prompt caching.

Long context pricing

Claude models have a large context window, which lets you send very long documents. Check whether the model you use has long context pricing, where requests above a certain size are charged at a different rate. Sending the full context window every time is rarely necessary, so trim documents to the sections that matter. At the time of writing, Claude 4.6 and later models, apart from Haiku 5.5, include the full 1 million token context window at standard rates. Claude Haiku 5.5 charges higher rates for any request whose prompt is over 100,000 tokens.

Tool use, web search and code execution

When Claude uses tools, the tool definitions are added to your input tokens on every request, and tool results come back as more input. Some built-in tools carry their own charges on top of tokens. Web search is usually charged per search, and the code execution tool and web fetch tool have their own pricing rules. At the time of writing, web search costs $10 per 1,000 searches plus tokens, and web fetch has no charge beyond tokens. Code execution is free when used with web search or web fetch. Otherwise each organisation gets 1,550 free hours a month, then pays $0.05 per hour per container. Agents that call many tools in a loop can use far more tokens than a single question and answer.

Fast mode, data residency and regional options

Some models offer a fast mode for quicker responses at premium pricing. Anthropic also offers data residency options, such as US only inference, which can carry a different rate from global routing. If you need data processed in a particular region, check the cost before you commit. At the time of writing, fast mode on Claude Opus 5.5 costs $8 per million input tokens and $40 per million output tokens, double the standard rate. US only inference on Claude 4.6 and later models costs 1.1 times the standard price.

Where to buy: the Claude Platform and cloud providers

You can use Claude directly through the Claude Platform and Claude Console, or through cloud providers such as Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry. Each has its own billing and data terms, and some set their own prices rather than using Anthropic's. Claude Platform on AWS and Claude in Microsoft Foundry bill at Anthropic's standard API rates through the cloud marketplace. Amazon Bedrock and Google Cloud set their own prices, and their regional endpoints cost 10% more than global ones.

For most small businesses, buying direct is the simplest place to start. If your business already runs on AWS, Google Cloud or Microsoft Azure, buying Claude through that provider can keep API billing on one invoice and inside your existing data agreements.

Claude Code pricing and how it relates to the API

Claude Code is Anthropic's coding tool. You can use it through a Claude subscription that includes Claude Code access, or connect it to the API and pay per token. Developers who use it heavily often find a subscription more predictable, while occasional use can be cheaper on the API. Claude Code pricing changes, so compare your expected usage against both options. At the time of writing, Claude Code is included in every paid Claude plan, from Pro at $20 a month ($17 a month billed annually) through Max, Ultra, Team and Enterprise, and it shares the plan's usage limits.

If developers use Claude Code with an API key, their usage appears on your API billing alongside your automations. Give them their own key or workspace so you can see what coding work costs separately.

Estimating your Claude API costs step by step

Here's a simple method you can use for any automation before you build it.

• Count the requests per month. For example, the number of emails, documents or chats.
• Measure the input tokens per request, including the system prompt, history and attachments. The token counting endpoint in the API or the usage figures in the Claude Console will give you real numbers.
• Measure the output tokens per request, including any thinking.
• Multiply input and output tokens by the current rates for your chosen model, then add them together.
• Apply caching to the repeated part of the input, and the Batch API rate if the job can wait.
• Add about 20% for retries, testing and growth.

Why real token usage beats guesswork

Estimates are useful, but your real token usage is what you pay for. Run the automation on a sample of real data for a week, then look at the usage reports. You'll often find the system prompt or chat history is bigger than you thought, which is good news, because those are the easiest parts to cache or trim.