Claude API for Beginners: A Practical Getting-Started Guide

The Claude API is a single, stateless HTTP endpoint — the Messages API — plus supporting endpoints for batches, files, and token counting. You send a list of messages (user and assistant turns), a model ID, and a token limit; Claude returns a structured response. Everything else — tool use, vision, streaming, structured outputs — is a feature of that one endpoint, not a separate product.

• Stateless: the API keeps no memory. You resend the conversation history on every call.
• Model-agnostic: swapping models is usually a one-string change.
• SDK-first: official libraries exist for Python (anthropic) and TypeScript (@anthropic-ai/sdk), with the same shapes in both.

Understanding this architecture early saves you from the most common beginner mistake: expecting the API to remember your last request.

Head to console.anthropic.com, create an account (or an organization for your team), and generate an API key from the settings area. Treat the key like a password: store it in an environment variable, never in source control.

Authentication is a single header — x-api-key — plus an anthropic-version header on raw HTTP calls. The SDKs read ANTHROPIC_API_KEY from your environment automatically, so most code never touches the key directly.

• Create separate keys per environment (dev, staging, production).
• Use workspaces to isolate team projects and spend.
• Set usage limits in the Console before you hand keys to a wider team.

Teams rolling this out org-wide often pair setup with our corporate Claude AI training so security and finance stakeholders are aligned from day one.

Here is the smallest useful request in Python. Install the SDK with pip install anthropic, export your key, then run:

import anthropic

client = anthropic.Anthropic()
msg = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Explain REST in one paragraph."}],
)
print(msg.content[0].text)

Three parameters are required: model, max_tokens (the output ceiling), and messages. The response contains a list of content blocks — check each block’s type rather than assuming plain text, because tool use and thinking can add other block types. The usage object on every response reports input and output tokens, which becomes your cost telemetry later.

As of 2026 the lineup has four working tiers, each addressed by a model ID string:

• Claude Haiku 4.5 (claude-haiku-4-5) — fastest and cheapest; classification, routing, high-volume simple tasks.
• Claude Sonnet 5 (claude-sonnet-5) — the balanced default for most production workloads, including coding.
• Claude Opus 4.8 (claude-opus-4-8) — hard reasoning and long-horizon agentic work.
• Claude Fable 5 (claude-fable-5) — the flagship for the most demanding reasoning tasks.

A sensible strategy: prototype on Sonnet, then benchmark down to Haiku where quality holds and up to Opus or Fable where it doesn’t. Our model guide covers the selection trade-offs in depth, and Anthropic’s pricing page has current rates per tier.

Once hello-world works, these five features cover most real applications:

• System prompts — a top-level system parameter that sets role, tone, and rules, separate from user messages.
• Multi-turn conversations — append each assistant reply to your messages array and resend; the first message must be a user turn.
• Streaming — server-sent events deliver tokens as they generate, essential for chat UX and long outputs.
• Vision — pass images as base64 or URL content blocks alongside text.
• Tool use — define functions with JSON schemas; Claude decides when to call them and your code executes and returns results.

Add prompt caching and batch processing when cost matters, and structured outputs when downstream code must parse the response.