tool_choice: auto, any, tool and none

Try 3 questions on this

Pick an answer to see why each option is right or wrong.

1. Your support service on Claude Sonnet 5 (claude-sonnet-5) sets tool_choice to {"type": "any"} and instructs the model in its system prompt to explain its reasoning in one sentence before calling a tool. The tool calls are correct, but the explanatory sentence never appears. What is the most likely cause?

   Right. In the game that’s Renown for your Tool Design, MCP and Server Tools building.

   Not that one: D is right. In the game your Tool Design, MCP and Server Tools building would lose a little Renown, and this question would come back later.

   Your answerA.The any mode strips text blocks from the response once generation completes.Nothing filters the response afterwards. The prefill means the turn begins at the tool call, so no preamble is ever generated.Your answerB.The any mode requires strict: true on each tool before natural-language output is allowed.Strict tool use guarantees schema conformance and is unrelated to whether text may precede a tool call. The two settings compose but neither gates the other.Your answerC.The any mode is ignored whenever the system prompt contradicts it, dropping text.tool_choice is enforced by the API rather than negotiated with the prompt. A contradicting instruction does not switch it off.The answerD.The any mode prefills the assistant turn, so no text precedes the tool_use block.With tool_choice set to any or to a named tool, the API prefills the assistant message to force a tool call. The model therefore emits no natural-language response or explanation before the tool_use blocks, however the prompt is written.

   Written from platform.claude.com/docs/en/agents-and-tools/tool-use/define-tools

2. Your service on Claude Opus 5 (claude-opus-5) caches a large system prompt and tool catalogue. For a small subset of requests it switches tool_choice from {"type": "auto"} to {"type": "any"} on an otherwise identical conversation, and monitoring shows those requests reading only part of the prefix from cache. What is the most likely root cause?

   Right. In the game that’s Renown for your Tool Design, MCP and Server Tools building.

   Not that one: C is right. In the game your Tool Design, MCP and Server Tools building would lose a little Renown, and this question would come back later.

   Your answerA.Changing tool_choice invalidates the cached tool definitions and the cached system prompt as well as the message blocks.Tool definitions and system prompts stay cached across a tool_choice change. Only message content has to be reprocessed.Your answerB.Changing tool_choice cannot affect caching at all, so the message content on those requests must differ.The interaction is documented: a change to tool_choice invalidates cached message blocks even when nothing else about the request differs.The answerC.Changing tool_choice invalidates the cached message blocks, but the tool definitions and system prompt stay cached.Changing tool_choice invalidates cached message blocks, so message content is reprocessed. Tool definitions and the system prompt remain cached, which is exactly the pattern the monitoring shows.Your answerD.Changing tool_choice invalidates the entire cached prefix, but only when the new value names one specific tool.Any change to the tool_choice parameter has this effect, not only a switch to a named tool.

   Written from platform.claude.com/docs/en/agents-and-tools/tool-use/define-tools

3. Your service shares one Claude Opus 5 (claude-opus-5) request builder across an agentic path and a high-volume classification path. The classification path never needs tools but still ships the full tool array with tool_choice: {"type": "none"}, and its prompts are short. You want the classification path to carry no automatic tool overhead and to stop paying repeatedly for tool definitions. What's the most effective way to meet these requirements at the lowest cost?

   Right. In the game that’s Renown for your Tool Design, MCP and Server Tools building.

   Not that one: A is right. In the game your Tool Design, MCP and Server Tools building would lose a little Renown, and this question would come back later.

   The answerA.Omit tools entirely on the classification path and keep tool_choice: {"type": "none"}, so no tool-use system prompt is injected at all.The tool-use system prompt is injected automatically whenever the request declares at least one tool: on Opus 5 that is 286 tokens with auto or none and 406 tokens with any or a named tool.Your answerB.Keep the tool array and rely on tool_choice: {"type": "none"}, which suppresses the injected tool-use system prompt on its own.tool_choice: {"type": "none"} tells the model not to call a tool, but the definitions are still rendered and the tool-use system prompt is still injected — 286 tokens on Opus 5 — on every classification request.Your answerC.Keep the tool array and put a cache_control breakpoint on the last tool definition so the short classification prompts read the catalogue from cache.Caching reduces the price of tokens that are still sent; it does not remove them. It also needs a cacheable prefix of at least 512 tokens on Opus 5, and short classification requests may never reach the threshold, leaving the write cost with no reads.Your answerD.Keep the tool array and set defer_loading: true on every tool, so no definition is materialized unless the model searches for one.defer_loading withholds a tool's schema until the model searches for it, which still requires a non-deferred search tool and at least one non-deferred tool — deferring everything returns 400. It is a catalogue-scaling feature, not a way to send no tools.

   Show the full explanation and doc quotes

   Overhead is zero only when no tools are declared and tool_choice is none. For a path that never calls a tool, omitting the array is both the cheapest and the simplest fix.

   Written from platform.claude.com/docs/en/agents-and-tools/tool-use/overview