Gemini API agents can now load tool schemas on demand, with no extra fee
Google's new preview feature keeps an agent's tool library on standby and loads full schemas only when a task needs them, at no separate charge.

Google has put Tool Retrieval into preview in the Gemini Interactions API. Agents can hold back full parameter schemas for most of their tools and load them only when a task calls for one, which Google says reduces input tokens and latency.
Two flags change how an agent carries its tools
The setup is small. A developer adds a tool_search tool to the request and sets defer_loading: true on the tools to hold back. According to Google, each held-back tool stays visible to the model through its name and short description, so the model knows the option exists, while the detailed parameter definitions are kept out of the prompt.

Once the model decides it needs one of those tools, it requests it by name; the API then fetches the full definition and the model issues the function call, all within a single request. Google’s example is a support agent with more than 50 deferred functions, where an issue_refund schema is pulled in only after a customer reports a damaged order. Each stage appears in interaction.steps as a tool_search_call, a tool_search_result and a function_call, which Google says helps developers debug tool names and descriptions.
Google describes the problem as one of scale. Once an agent connects to dozens or hundreds of tools, sending every schema on every request inflates input tokens, adds latency and, per Google, lowers tool-selection accuracy as the candidate list grows.

Server-side search by default, client-side for custom registries
By default Gemini runs the search on the server, so there is no index to build or host. Developers can keep frequently used tools loaded up front and defer only the long tail.
Teams that already maintain their own tool registry, embedding index or per-user permission filter can declare tool_search with "execution": "client" and define their own search parameters. That path adds one round trip per search, because the app runs the lookup. Google notes that tools do not have to be declared upfront in this mode, so an app can expose new tools mid-conversation, for example when a user connects an integration.

The feature also reaches remote MCP servers. Setting defer_loading: true on an mcp_server tool defers every tool the server exposes, so a large server can be connected without listing its tools one by one. Google also says the design works with prefix caching: deferred schemas stay out of the start of the prompt and loaded ones are added later, so the cached prefix does not change between turns.
Tool Retrieval is in preview in the Interactions API and works with Gemini 3.6 and later models, including all Gemini 4 models, according to Google.
No extra fee, but the savings depend on your tool mix
Google says the feature carries no extra fee; usage is metered at the ordinary per-token prices of whichever model is called. Enabling tool_search adds a small fixed amount of input tokens, and any schema the model pulls in keeps counting as input for the remainder of that conversation, exactly as if the tool had been declared from the start.
That makes the economics a matter of ratio. The saving comes from the schemas an agent never loads outweighing the ones it does, which Google calls the usual case for large libraries and tells developers to measure against their own traffic. An agent with a handful of tools that are all used on most turns has little to gain, and the fixed search overhead is then pure cost.
The size of that saving depends on the model’s rates. Our earlier report on Gemini 4 Hydrogen listed $1.50 per million input tokens and $7.50 per million output tokens, and $2.00 and $10.00 for Gemini 4 Carbon. At those input rates, every 100,000 tokens of schemas that an agent no longer resends cost between 15 and 20 cents per request, a figure that compounds across long, multi-turn tasks. That is simple arithmetic on listed prices, not a result Google has published.
The feature also fits a wider push on per-task cost. As we explained in our guide to intelligence indexes and cost per task, Artificial Analysis added per-task cost and token metrics to its index in version 4.1, published in June 2026, because per-token price alone says little about what a finished job costs. Tool Retrieval attacks the input side of that bill rather than the model’s price.
What the announcement leaves unproven
Google’s post gives no measured token savings, latency figures or accuracy numbers, so the claims of lower input counts and better tool selection are the company’s own and have not been independently checked. It does not say when the preview will reach general availability, or whether it will arrive in other Gemini API surfaces beyond the Interactions API. The client-side route has a cost Google states plainly, an extra round trip per search, and the post does not quantify its latency.
For developers, the practical step is to run an existing agent with its rarely used tools deferred and compare billed input tokens and tool-call accuracy against the current setup. The documentation, a model comparison page and a Discord channel for feedback are linked from Google’s announcement.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.
You may also like

Perplexity’s Orchestrator API puts 85+ models behind one key and no platform fee

Google’s Decisions API scores 1,000 questions at once for $0.10 per million tokens

Gemini can now run your Mac apps while you keep working
