The catalog stays visible while schemas wait
In Google’s example, developers add tool_search and mark selected functions with defer_loading. The model sees their names and descriptions, then requests the detailed schema when needed. Frequently used functions can remain loaded. The launch supports Gemini 3.6 and later models, including Gemini 4.
Google says there is no separate retrieval fee: normal model-token charges apply. This is a change to how capabilities reach the model, not a new model release or a fine-tuning method. The feature does not replace writing and maintaining the underlying functions.
Fewer initial schemas do not guarantee a cheaper task
The retrieval guide explains that a server-side search triggers another model pass with the full context and newly loaded schemas. That prompt contributes to total_tool_use_tokens. Multiple searches can therefore add costs beyond the initial request, even when the catalog itself is compact.
Retrieved tools remain in context and cannot be unloaded during the conversation. The guide also limits retrieval to Interactions rather than generateContent. These constraints matter when comparing an existing integration with a preview, especially for long sessions that eventually use many tools.
Function execution and conversation state remain application work
Google’s function-calling guide separates generating a function call from executing it: application code receives the name and arguments, runs the operation and returns a result. Discovering a function is therefore not evidence that the operation has completed.
The same guide requires sending tool declarations again on subsequent requests, including with previous_interaction_id. In stateless mode, returned tool-search steps and their signatures must be carried forward unchanged. Remote MCP support has additional transport constraints: Streamable HTTP is supported, while SSE servers are not.
Evaluate the complete workflow, including retrieval misses
Our assessment: begin with a small, representative task set and compare full-schema loading with deferred loading. Record the correct final result, time to completion, total billed tokens and the number of searches. A catalog containing similarly named operations is useful for testing whether descriptions actually distinguish them.
Include cases where the required function is unavailable and where a user lacks permission for the requested operation. A tool lookup should not expand that user’s rights. For example, finding a refund function and successfully authorizing a refund are separate checks. We have not run a comparative benchmark, so this report makes no percentage claim about savings or accuracy.

Discussion.
Your perspective matters. Sign in, pick a nickname and choose your cyber identity.
Loading discussion…
Your account name, email and photo stay off your comments.