Claude API Guide: Integrate Anthropic's AI in Your App

Getting API Access

Start at console.anthropic.com. Create an account, verify your email, and navigate to the API Keys section. Create a new API key — this key authenticates all your API requests.

Store the key securely: in environment variables (never in code), in a secrets manager for production deployments, and never in version-controlled files. An exposed API key is a significant security risk as it allows unlimited API usage billed to your account.

Anthropic recommends setting usage limits on new keys during development to prevent runaway costs from bugs or infinite loops. Set a monthly spending limit appropriate to your expected usage.

The Messages API

All Claude interactions use the Messages API (POST /v1/messages). The request structure:

Model: The Claude variant to use. Current production models:

• claude-haiku-4-5-20251001 — Fast, cheapest, for high-volume simple tasks
• claude-sonnet-4-6 — Balanced capability and cost, most common choice
• claude-opus-4-8 — Maximum capability for complex tasks

max_tokens: The maximum number of tokens in the response. Set this conservatively — you're charged for output tokens. For classification tasks where the output is small, set this to 200-500. For long-form generation, 2000-4096.

system: The system prompt that defines Claude's role, behaviour, and output format. This is the most important element of your API request.

messages: The conversation history as an array of role/content pairs. For automation (single-turn), this is just one user message. For multi-turn conversations, include the full history.

Model Selection Strategy

Model selection is the most important cost-performance decision in your Claude implementation.

Haiku for: Simple classification, yes/no decisions, short data extraction, high-volume tasks where speed and cost matter most. At $0.25/$1.25 per 1M tokens (input/output), Haiku is the right choice for tasks that don't require sophisticated reasoning.

Sonnet for: Complex analysis, long-form generation, multi-step reasoning, customer-facing responses where quality matters. At ~$3/$15 per 1M tokens, this is appropriate for tasks where quality justifies the cost.

Opus for: The most complex tasks — nuanced legal analysis, advanced research synthesis, complex code generation. Use selectively for the highest-value tasks in your workflow.

A mixed model strategy in an automation workflow might use: Haiku to classify incoming support tickets (high volume), Sonnet to draft responses for standard issues (moderate volume, quality matters), and Opus for complex escalation analysis (low volume, maximum quality needed).

Prompt Engineering Essentials

Be specific and explicit: Claude follows instructions literally. If you want a specific output format, describe it exactly. If there are conditions or exceptions, state them explicitly. Don't rely on implication.

Structured output: For automation, always request structured output. Specify the exact JSON schema you expect. Claude will produce it consistently when clearly specified:

Role definition: The system prompt should clearly define Claude's role, what it knows, and what it's responsible for. A well-defined role improves consistency.

Handle edge cases explicitly: Tell Claude what to do when the input is ambiguous, empty, in an unexpected format, or doesn't match expected patterns. If you don't, Claude will make a reasonable guess — which may not match what your workflow expects.

Error Handling and Retry Logic

The Anthropic API returns standard HTTP status codes. Handle these explicitly:

400 Bad Request: Usually a malformed request or prompt exceeds max tokens. Check your prompt length and request format.

401 Unauthorized: Invalid API key. Check the key is correct and not expired.

429 Rate Limited: You've exceeded the rate limit. Implement exponential backoff: wait 1 second, retry; if still rate limited, wait 2 seconds; then 4 seconds; then 8 seconds. Stop after 5 retries.

500/529 Server Error: Anthropic server issues. Retry with exponential backoff.

For production automations, implement retry logic for 429 and 5xx errors. Most transient errors resolve within 30 seconds.

Cost Optimisation

Minimise prompt tokens: Every token in your system prompt and user message is charged. Audit your prompts for unnecessary verbosity. Remove instructions that don't change Claude's behaviour. Shorter prompts = lower cost = lower latency.

Right-size models: The most common cost mistake is using Sonnet or Opus for tasks where Haiku produces acceptable results. Test classification and extraction tasks with Haiku before defaulting to larger models.

Cache system prompts: Anthropic's prompt caching feature allows caching long system prompts, reducing input token costs by up to 90% for high-frequency tasks with stable system prompts.

Streaming for UX, not cost: Streaming returns tokens as they're generated (better UX for interactive applications) but doesn't reduce cost.

Recommended Tools

• Anthropic API — The API itself
• Claude.ai — Prompt testing and iteration
• n8n — Integrate Claude API into automation workflows
• Make.com — Visual Claude API integration
• Postman — API testing during development