The Claude API: pricing, models and first steps.

Subscription or API: two different uses.

Claude can be used in two ways, billed separately:

• A subscription (Free, Pro, Max, Team, Enterprise) gives access to the Claude app, on the web, desktop and mobile. It is a tool for people. Prices are covered in Claude Pricing 2026.
• The API lets your software, automations and agents call Claude directly. You pay per use, per million tokens, with no subscription.

A token is a small piece of text. According to Anthropic, one token is roughly 4 characters or 0.75 words in English; the exact count varies by language.

The API is what you use as soon as you want Claude to work inside a process: sort emails, extract invoice data, draft replies, run an AI agent.

The available models.

As of September 27, 2026, Anthropic's documentation lists four current models:

Anthropic's advice is simple: when in doubt, start with Claude Opus 5.5, and move to Fable 5.1 if your tests show Opus 5.5 falls short. All current models read text and images and can call tools.

Claude API pricing.

Here are the prices listed by Anthropic on September 27, 2026, in US dollars, per million tokens:

Three details. The 1 million token context is billed at the standard rate: a 900,000-token request costs the same per token as a 9,000-token one. Sonnet 5's price ($2 and $10), first announced as introductory, is now the standard price. And web search through the API costs $10 per 1,000 searches, on top of tokens.

A worked example: classifying 1,000 emails with Haiku 4.5, reading 2,000 tokens and writing 300 tokens per email, means 2 million input tokens ($2) and 300,000 output tokens ($1.50). That is $3.50 for the 1,000 emails, and half that with batch processing.

Paying less: batch, cache, right model.

Three levers, which stack according to the documentation:

• Batch processing. Requests that do not need an immediate answer (overnight jobs, bulk analysis) cost 50% less.
• Caching. When every request starts with the same long prefix (instructions, reference documents), that prefix can be cached. Writing costs a bit more (1.25x the input price for 5 minutes, 2x for 1 hour), but each read costs only 10% of the input price, even less on Opus 5.5 and Fable 5.1.
• The right model for each task. A simple sort does not need the most capable model. In a well-designed system, Haiku handles the volume and Opus steps in on the hard cases.

Where your data goes.

By default, Claude API inference may run in any available geography, for performance and availability. According to the documentation, two options let you frame this:

• On Anthropic's API, the inference_geo parameter restricts inference to the United States, with a 10% surcharge on tokens. It is the only geographic restriction offered on the direct API today.
• On Amazon Bedrock and Google Cloud, Claude is available through regional endpoints, which guarantee data is routed through a specific geographic area, for 10% more than global endpoints. Available regions depend on the provider and the model.

For a company with strict data location requirements, regional cloud endpoints are the first route to look at, under the cloud provider's contract. And for tasks where data must stay in a given region, other models remain an option: see Which AI to Choose in 2026.

Making a first call.

Four steps to test the API:

1. Create an account on the Claude Console (platform.claude.com).
2. Generate an API key in the settings, and keep it like a password.
3. Add credit (credit card for standard accounts).
4. Send a first request, for example from a terminal:

curl https://api.anthropic.com/v1/messages \
  -H "content-type: application/json" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-opus-5-5",
    "max_tokens": 1000,
    "messages": [{"role": "user", "content": "Summarize this contract in five points."}]
  }'

No code needed to go further: automation tools like n8n or Make call the Claude API from their own modules. The step-by-step method to build an agent is in How to Build an AI Agent with n8n, and for coding, see Claude Code.

Claude API: what people ask us.