AI Advanced: Claude SDK & Anthropic APIs Practice Questions

Statement B is correct: cache hits are charged at a significantly reduced input token rate (roughly 10% of the standard input price), making repeated identical prefixes much cheaper. Statement D is correct: the minimum cacheable prefix for Claude 3 models is 1,024 tokens; shorter blocks will not be cached. — Statement A is wrong on the TTL: the ephemeral cache TTL is 5 minutes, but it is reset (extended) on each cache hit, not simply refreshed — however more importantly, the 5-minute TTL detail itself is accurate but 'refreshed on every cache hit' overstates the behavior; the TTL resets to 5 minutes from the last hit, which is a nuanced but relevant distinction making this option partially misleading and not a clean 'accurate' statement compared to B and D. Statement C is wrong: prompt caching is available across multiple Claude 3 model tiers, not only Haiku. Statement E is wrong: caches are scoped per API key and model, but different API keys do NOT share a cache namespace. Statement F conflates input caching with output billing; cached tokens still count toward context window limits but the statement about output token billing is a non-sequitur — output tokens are priced separately regardless.

Option C is correct: the Anthropic API requires the full conversation history in each request. After Claude returns an assistant message containing a tool_use block, you must include that assistant message in the history and then add a new user message whose content is a tool_result block that references the tool_use_id from the assistant's tool call. Option A is incorrect; there is no "tool" role in the Anthropic API — tool results are submitted as user role messages with a tool_result content block type. Option B is incorrect about the need to re-include the assistant message: the Anthropic API is stateless, and all prior messages (including the assistant's tool_use message) must be sent with every request. Option D is incorrect; tool_result blocks belong in a user role message, not an assistant role message.

Option B is correct: the Anthropic Python SDK provides a client.messages.stream() context manager that yields a stream object; iterating over stream.text_stream yields successive text delta strings, which is the idiomatic high-level streaming API. Option A is incorrect because stream=True is the OpenAI SDK pattern; the Anthropic SDK uses client.messages.stream() as a context manager, not a stream=True flag on messages.create. Option D is incorrect because create_stream is not a method in the Anthropic SDK; it is a fabricated name. Option C is incorrect because stream_create and stream.done() / stream.next_chunk() are not part of the Anthropic SDK's interface.

Claude Desktop's MCP configuration uses a mcpServers object where each entry specifies a command (the executable) and args (an array of arguments). For a Node.js script, command should be "node" and args should contain the script path. Option D describes an HTTP-based transport format; while some MCP transports use HTTP, Claude Desktop's stdio-based local server config requires command/args, not a URL. Option C uses a tools array with an executable key—this schema does not match Claude Desktop's MCP configuration format at all. Option B uses script and runtime keys, which are not valid fields in the Claude Desktop MCP server configuration spec.

Option B is correct: cache_control is placed directly inside an individual content block object (when content is an array of block objects), not at the message level. The breakpoint is a property of the specific content block where the cache boundary should be set. Option D is wrong because placing cache_control at the message level (alongside role and content) is not the correct structure; it must be inside a content block object when content is structured as an array. Option A is wrong because there is no cache_marker content block type in the Anthropic API; cache_control is added directly to existing content block types like text or image. Option C is wrong because cache: true is not a valid field; the correct field name and structure is cache_control: {"type": "ephemeral"}.