claude-api

Use client.messages.stream() as a context manager and iterate over stream.text_stream to print tokens as they arrive. In TypeScript, iterate the stream with for await and handle content_block_delta events containing text_delta payloads.

Define tools with name, description, and input_schema, pass them to messages.create, then inspect response content blocks for tool_use types. Execute the tool locally and send results back in a tool_result content block to continue the conversation.

Use Sonnet 4 (claude-sonnet-4-0) for most development tasks, Opus 4.1 for complex reasoning, and Haiku 3.5 for high-volume cost-sensitive workloads. For production, prefer pinned snapshot IDs over aliases.

Yes, add cache_control with type ephemeral to system prompt blocks to cache large contexts. Cached tokens can reduce costs by up to 90 percent, and usage is reported via cache_read_input_tokens and cache_creation_input_tokens.

Catch RateLimitError from the anthropic package and back off before retrying, typically waiting around 60 seconds. Also handle APIConnectionError for network issues and generic APIError for other status codes.

Use the Batches API for non-time-sensitive bulk processing, which offers a 50 percent cost reduction. Submit requests with custom_id values, poll the batch status until processing ends, then iterate over results.