claude-api

Install the anthropic package with pip, create an anthropic.Anthropic() client that reads ANTHROPIC_API_KEY from the environment, and call client.messages.create with a model, max_tokens, and messages array. The response text is available at message.content[0].text.

In Python, use client.messages.stream as a context manager and iterate over stream.text_stream. In TypeScript, use client.messages.stream and loop over events, printing text from content_block_delta events with text_delta types.

Default to Sonnet 4.6 for balanced coding and development tasks. Use Opus 4.6 for complex reasoning and architecture work, and Haiku 4.5 for fast, high-volume, cost-sensitive tasks like classification and extraction.

Yes, define tools with a name, description, and JSON input_schema, pass them in the tools parameter, and handle tool_use content blocks in the response. Execute the tool locally and send results back as tool_result messages to continue the conversation.

Use prompt caching for repeated system prompts (up to 90% savings on cached tokens), the Batches API for non-urgent bulk work (50% savings), and Haiku instead of Sonnet for simple tasks (~75% savings). Setting shorter max_tokens also reduces spend.

Catch RateLimitError from the anthropic package and back off before retrying, typically sleeping around 60 seconds. Also handle APIConnectionError for network issues with retry backoff and APIError for general failures with status codes.