Claude API Config Builder

Current Claude models: claude-opus-4-6 (most capable), claude-sonnet-4-6 (balanced speed/quality), claude-haiku-4-5-20251001 (fastest/cheapest). Always use full model IDs to avoid unexpected version changes. Claude 3 models (Opus/Sonnet/Haiku 3.x) are still available for backward compatibility.

The system prompt sets Claude's behavior, persona, and constraints before the conversation begins. Human/assistant turns represent the actual dialogue. For best results, put instructions, context, and formatting requirements in the system prompt rather than repeating them in each human message.

Set stream: true in your request. The API returns server-sent events (SSE). Each chunk contains a delta with partial text. In Node.js, use the SDK stream() method; in Python, use with client.messages.stream() as stream. Streaming reduces time-to-first-token significantly for long outputs.

Define tools with clear names, descriptions, and JSON Schema for parameters. Claude decides when to call tools based on context. For forced tool use, set tool_choice type to auto or tool. Always handle tool_use stop_reason and inject tool results in the next human turn.

Implement exponential backoff on 429 errors. Use the x-ratelimit-* response headers to monitor limits. For high throughput, use batching (Anthropic Batch API) for non-realtime workloads. Enterprise tier has higher limits. Haiku is the best choice for high-volume use cases.