
Why it mattersA team with many tools pays input tokens for every one on every call. Deferring load moves that cost to the tool the model asks for by name.
An agent that can call 50 tools pays input tokens for 50 schemas on every turn. The model may use one of them or none. Google's new tool retrieval in the Gemini API asks the model to pick a tool by name. The API then loads that tool's schema in the same request. The schemas the model never calls are never billed.
Google announced the preview on November 19. The feature works with Gemini 3.6 and later, including every Gemini 4 model. It runs in the Interactions API.
How it works
A developer adds the tool_search tool to the call and sets defer_loading: true on each tool they want to hold back. The model still sees every tool's name and description. It does not see the deferred tools' parameter schemas. When the model decides it needs one, it asks for it by name. The API loads that tool's full schema and the model calls the function. It all happens inside one request.
Google says there is no index to build. The feature is server-side. A developer adds the deferred flag in code, runs the call, and the API does the rest.
What it costs
There is no separate charge for tool retrieval. Standard input and output token rates apply. The tool_search tool adds a small fixed token block to the call. A tool's full schema counts as input tokens for the rest of the conversation once it has been loaded.
Google writes that the feature lets a developer "reduce latency and input token costs for complex tasks". The post tells developers to measure the result "against your own traffic." It does not quote a specific percentage. A team should run the same workload with the flag off and with it on. Then compare the input-token count and the time to first byte.
Where it fits
The post lists two uses. One is a long list of client-side tools: the developer keeps their own registry and marks most of it deferred. The other is a remote MCP server: the model sees the names on the registry and asks for the schema of the one tool it needs.
The practical change is how a team on Gemini runs an agent against a long tool list. Until today, every tool on the call burned tokens on every turn, even the tools the model never called. The deferred flag moves that cost to the tool the model picks. A developer who keeps the full list in the call for every turn pays the full cost. A developer who defers the rest pays for one.
The feature is in preview in the Interactions API. A team that already runs one agent on a long tool list can test it on one workload today. Keep that first measurement as the number to beat.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


