What actually shipped

Sending every tool schema up front is the common setup Google's post describes, and the post states its cost plainly: it inflates input token counts, adds latency, and lowers tool-selection accuracy as the candidate list grows.

Tool Retrieval splits the declaration in two. The name and a short description stay in the prompt. The schema stays out. When the model decides it needs a tool, it asks for it by name inside the same request, the API loads the full schema, and the model emits the function call. The search, load and call steps are recorded in interaction.steps. Google's worked example defers more than 50 functions and loads one, a refund function, on demand.

The same flag works on a remote MCP server: set defer_loading: true on an mcp_server tool and every tool that server exposes loads on demand. Developers who run their own registry can choose client-side retrieval instead, at the cost of one extra round trip per search. Deferred schemas also stay out of the start of the prompt, which keeps the cached prefix stable from turn to turn.

Google is not the first to ship the pattern. PPC Land's comparison notes similar deferred-loading designs released earlier by Anthropic, in November 2025, and OpenAI, in March 2026.

Where the saving stops

The pricing section is the part to read slowly. Tool Retrieval has no separate charge, which sounds like a free efficiency gain. Then the conditions arrive. Turning on tool_search adds a small, fixed block of input tokens to the request. Each schema the model does load is billed as input from the point it loads through the rest of the conversation, the same as a tool declared up front. The saving comes from the schemas never loaded, and from the turns before a schema loads; how much that is worth depends on schema size, when and how often schemas load, the fixed overhead, and cache behavior.

Google says this in its own words: you save tokens when the schemas you never load outweigh the ones you do, which it calls the usual case for large libraries, followed by the instruction to measure it against your own traffic. That is an honest sentence. It is also the whole of the evidence offered. There is no percentage, no benchmark run, no before-and-after token count for a stated workload. An agent that touches most of its tools in a long session could save little, and pay for the search steps on top.

There is a second, quieter limit. Until a schema loads, the model works from names and descriptions alone, so clear names and descriptions carry more weight than they did when every schema was visible: retrieval has to find the right tool before its parameters are in context at all. Google's debugging advice, to inspect interaction.steps when tool names or descriptions mislead the model, points at the same dependence.

What to do with it

If your agent exposes dozens of tools and most requests use a handful, turn it on in preview and compare billed input tokens, latency and tool-selection accuracy on representative conversations before depending on it. Keep frequently used tools loaded up front, as Google's example does, and defer the long tail. If your tool set is small and its schemas are short, measure before assuming there is anything to gain; the search step adds its own fixed block of tokens either way.

Sources

  1. [1] Google, "Gemini API tool retrieval for efficient agent scaling," November 19, 2026 (mechanics, model support, MCP support and full pricing terms)Read source
  2. [2] PPC Land, "Google shrinks idle tools in Gemini agents to about 26 tokens each," November 2026 (context on similar earlier deferred-loading designs from other providers)Read source