Tool to retrieve details of a specific model by its ID. Use after confirming the model ID is valid.

Tool to list available models. Use when you need to see which models are available before selection. Always source model IDs from this tool rather than hard-coding them; stale or outdated model IDs passed to CREATE_MESSAGE will cause a not_found_error.

Create messages with prompt caching to reduce costs by up to 90% and latency by up to 85% on repeated API calls. Mark static content (system instructions, documents, context) for reuse with cache_control. Cache reads cost 0.1x base price, writes cost 1.25x (5m TTL) or 2x (1h TTL). Default 5-minute TTL: cache_control: {"type": "ephemeral"}. One-hour TTL: cache_control: {"type": "ephemeral", "ttl": 3600}. Invalid or mismatched cache_control fields fail silently with no error. Up to 4 cache breakpoints supported. For conversations, cache message history by marking final user message. Minimum tokens: Opus 4.5/Haiku 4.5 need 4096, Sonnet 4.5/4 need 1024, Haiku 3.5/3 need 2048. Monitor cache_read_input_tokens and cache_creation_input_tokens in usage response to track effectiveness. Best for: RAG with large documents, agentic tools, long conversations, coding assistants, reference analysis.