AI APIs have unique design challenges due to LLM latency and non-determinism. Key design principles: 1) Always support streaming: AI responses take 5-30 seconds - streaming Server-Sent Events (SSE) or WebSocket provides immediate feedback to users. FastAPI example: use StreamingResponse with async generator yielding tokens. 2) Implement request idempotency: LLM failures are common - clients must safely retry. Accept Idempotency-Key header, cache responses keyed by idempotency key. Return same response for duplicate requests. 3) Tiered rate limiting: separate limits for free/paid tiers, implement token-based limits (not just request-based). 10,000 tokens per minute is more meaningful than 100 requests per minute for LLM APIs. 4) Handle LLM errors gracefully: implement circuit breaker pattern for upstream LLM API failures, fallback model strategies, proper error codes distinguishing temporary (503) from permanent (400) failures. 5) Request queuing for async workloads: accept request, return job ID immediately, process asynchronously, provide status polling endpoint. Good for batch analysis, document processing. 6) Semantic versioning for prompts: breaking prompt changes (different output format, different behavior) require API version bump. Non-breaking improvements can be rolled out transparently. 7) Cost attribution: inject customer/feature identifiers in LLM API calls via metadata for per-customer cost tracking.