Key takeaways
- Google announced Tool Retrieval preview on November 19, 2026. It requires the v1beta Interactions API; generateContent does not support it.
- Add tool_search and defer_loading: true to load function parameter schemas only when needed. Your application still executes ordinary function calls.
- No separate feature charge does not mean free search: the documentation bills internal search prompt tokens in total_tool_use_tokens at the input rate.
- Use it first for a large library with sparse tool use. Compare full-task cost and completion quality against static loading before adopting it broadly.
What Google launched
Google announced Tool Retrieval in preview on November 19, 2026, letting Gemini agents load function parameter schemas on demand instead of keeping an entire tool library in the initial model context. It is available through the Interactions API, using v1beta. The existing generateContent method does not support it. This is an API capability update, not a new model or a downloadable release.
The linked documentation explicitly lists gemini-4-carbon, gemini-4-neon, gemini-4-hydrogen, gemini-4-argon, gemini-3.8-flash and gemini-3.6-flash. Google’s launch post describes support more broadly as Gemini 3.6 and later, including all Gemini 4 models. Use the documentation’s explicit list when planning an integration rather than assuming every endpoint is covered. AccessAllGPT has not verified access in an authenticated account.
Schemas arrive when the agent needs them
For server-side retrieval, add {"type":"tool_search"} to the tools array and set defer_loading: true on the functions you want to hold back. The model still sees names and compact descriptions, but not the full parameter schemas. When it selects a function, the API retrieves its schema and returns a function call within the same request. There is no separate search index for the developer to host.
Google exposes tool_search_call, tool_search_result and function_call in interaction.steps. For ordinary functions, your application runs the function and returns a function_result; schema retrieval does not execute your business logic for you. Frequently used functions can remain loaded while the long tail is deferred. Google also advises leaving tools named explicitly in system instructions loaded, because the model may otherwise call them without searching first.
For a private registry or per-user tool catalog, execution: "client" delegates search to your application. That lets you supply tools dynamically, but adds one round trip per search. The two retrieval modes cannot be combined in a single request. Our analysis: server-side retrieval is the simpler starting point for a stable catalog; client-side retrieval is useful when the application needs to control which schemas a user can discover. Neither replaces authorization when a function runs.
No separate charge is not the same as free search
Google says Tool Retrieval has no separate feature charge: the selected model’s standard token rates apply. Its launch post emphasizes the smaller initial prompt and says savings depend on which schemas are loaded. The linked documentation supplies a material additional detail: each server-side search triggers an extra model call inside the request, with the full context plus loaded schemas. Those prompt tokens appear in total_tool_use_tokens and are billed at the input rate.
That means one HTTP request is not necessarily one model call or one prompt’s worth of billing. Deferred names and short descriptions still contribute to total_input_tokens, and schemas become billable once loaded. Tools that have been loaded remain in context for the conversation and cannot be unloaded. The documentation explicitly warns that static loading can cost less for small libraries or very long contexts that search on every turn.
Retrieved schemas are appended rather than rewriting the prompt prefix, which can preserve implicit caching. But deferral can also shrink a medium prompt below the model’s minimum cacheable size. There is no universal savings percentage in these materials. We did not measure a bill, and we do not interpret the vendor’s reduced-latency and improved-selection claims as independently established results.
Multi-turn clients need a careful upgrade
The documentation requires the full tools array on every request, including follow-ups that return function_result, because tool declarations are not persisted across turns. With previous_interaction_id, the API stores prior retrieval steps, but that does not remove the requirement to resend the tool declarations.
For stateless conversations, preserve every model-generated step in order, including retrieval steps and encrypted signature fields. Google documents a 400 error for omitted tool_search_call or missing signatures; omitting tool_search_result instead drops loaded schemas from context. Use an SDK that understands these step types rather than serializing only the final function call.
Other preview limits matter for test coverage: only one tool_search entry is allowed per request, the model can search multiple times in a turn, and a search with no matching deferred tool yields text rather than a structured error. These are documented behaviors, not failures observed in an AccessAllGPT trial.
Why it matters and what to try next
Large agents often expose many integrations while using only a handful for a particular task. Deferring schemas gives builders a managed alternative to writing their own selection layer, while retaining inspectable retrieval steps. The buying decision is whether that smaller active context improves completed-task economics and correct tool selection enough to justify a preview-only API dependency.
Start with synthetic tasks and a read-only tool catalog. Compare static loading against server-side deferral on identical tasks, recording correct function selection, valid arguments, completed tasks, end-to-end latency and all billed token categories, including total_tool_use_tokens. Repeat over multi-turn sessions so schemas accumulating in context and cache eligibility are visible. This is a proposed evaluation, not a test we performed. Keep a static-loading fallback until your own results justify the change.
Choose a tool-loading mode
A proposed integration decision, based on Google’s documented trade-offs rather than measured AccessAllGPT performance.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Keep static loading as the baseline. Search overhead can outweigh schema savings; compare whole-task bills, not initial prompt length.
Trial server-side retrieval. Keep frequent and explicitly named functions loaded; defer infrequently used schemas and inspect retrieval steps.
Consider client-side retrieval with application-owned filtering. Budget the extra round trip and enforce permissions again at execution.
Primary sources
Browse the publication-wide evidence index →
- Introducing Tool Retrieval in the Gemini APIGoogle · Reviewed: November 19 date; preview launch; how it works; server-side and client-side examples; pricing and supported-model statements · Retrieved · Supports: First-party announcement date, availability description, intended schema-loading workflow and vendor efficiency claims. Does not establish independently measured savings, latency or accuracy.
- Gemini API documentation: Tool retrievalGoogle AI Studio · Reviewed: v1beta and Interactions-only requirement; server-side response; choosing a mode; configuration; multi-turn and stateless replay; pricing and total_tool_use_tokens; supported models; best practices and limitations · Retrieved · Supports: Explicit model IDs, API restrictions, internal-search prompt billing, caching caveats, retained schemas, replay requirements and preview limitations. Documentation examples are not AccessAllGPT API results.
Limitations
Public-source briefing only. No authenticated Gemini access, SDK execution, API request, independent benchmark, latency measurement or invoice verification. Google’s efficiency claims remain vendor claims. Preview behavior, model support and billing details can change; no numerical savings claim is made.
Disclosures
Public-source reporting, not a sponsored test. AccessAllGPT Research is operated by NeuralArc and is independent of Google. Google did not review or endorse this article; publication-wide relationships appear on the disclosures page.



