ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI information
Gemini API Tool Retrieval Enters Preview: Load Hundreds of Tools on Demand, Spend Fewer Tokens and Pick More Accurately

Gemini API Tool Retrieval Enters Preview: Load Hundreds of Tools on Demand, Spend Fewer Tokens and Pick More Accurately

AI information • Admin • • 20 views

On November 19, 2026, Google announced on its official blog that Tool Retrieval is coming to the Gemini API in preview, for the Interactions API. The more tools an agent connects, the longer the tool parameter descriptions it has to carry on every request — and token spend, latency, and the odds of picking the wrong tool all climb with it. The new approach is simple: the tool library stays on standby, and parameter schemas load only when the task actually needs them.

To turn it on, add a tool_search tool to the request and mark the tools you want to hold back with defer_loading: true. The model still sees every tool's name and description, so it knows what is available, but the parameter schemas stay out of context; when a tool is needed, the model asks for it by name, and the API completes the search, the schema load, and the function call inside the same request, with every step inspectable in interaction.steps. A remote MCP server can also defer its whole tool set without listing each tool it exposes, and developers who maintain their own tool registry can set tool_search to run client-side, connecting new tools mid-conversation. Deferred schemas never enter the start of the prompt, so the feature also works well with prefix caching.

On pricing, there is no separate charge: input and output tokens are billed at the standard rates for the model you call, tool_search itself takes only a small fixed block of input tokens, and each schema the model loads counts as input for the rest of the conversation. Whether you save money comes down to the math: you come out ahead when the tokens saved on schemas you never load outweigh the ones you do, and the bigger the library, the better the deal. It supports Gemini 3.6 and later, including all Gemini 4 models. This continues a recent Gemini API theme of spending tokens only where they matter: it recently added support for reading just the needed parts of Office files, too.

The official docs already carry a full setup guide. For developers whose agent libraries run into the dozens, now is a good time to test it on your own traffic: keep high-frequency tools loaded up front, defer the whole long tail, then compare the token bill and the tool-selection error rate before rolling it out everywhere.

Recommended Tools

More