Grounding with Google Search pricing: per query, not prompt

Grounding with Google Search pricing for Gemini 3 and newer models, Gemini 4 Argon included, is 5,000 free search queries a month and then $14 per 1,000 (Gemini API pricing, read 2026-10-15). Google bills each search the model decides to run, not each prompt, so one question that triggers four searches counts four times.

Many pages still quote $35 per 1,000 grounded prompts. That is the rule for Gemini 2.5 and older models, and Google Cloud still lists it for them (Agent Platform pricing, read 2026-10-15). This post sets out what grounding costs on each Google platform, how a search is counted, what a real volume costs, and when an outside search tool is the better buy.

Last checked 2026-10-15, against Google's docs and pricing pages and looot's live catalog.

Grounding with Google Search pricing, by model and platform

Gemini 3 and newer models, on both the Gemini API and Google Cloud's Agent Platform, share one rule: a free allowance of 5,000 queries a month, then a price per 1,000 queries. Older models on Agent Platform still pay per prompt, with a free allowance counted per day instead.

Grounding is a paid-tier feature for Argon, because Argon itself has no free tier. On Agent Platform, Google adds that input tokens provided by grounding are not charged. The Gemini API pricing page does not say the same, so check your own bill before you count on it.

How is a grounded search counted?

The model decides how many searches to run for each prompt, and each one is a billed query. Google's grounding docs (read 2026-10-15) say that for Gemini 3 and newer "your project is billed for each search query that the model decides to execute", while Gemini 2.5 and older are billed per prompt.

You turn grounding on with one tool in an Interactions API call:

from google import genai

client = genai.Client()
interaction = client.interactions.create(
    model="gemini-4-argon",
    input="Who won the euro 2024?",
    tools=[{"type": "google_search"}],
)

The response holds a `google_search_call` step with the queries the model ran, a `google_search_result` step with search suggestions to show in your interface, and the answer with `url_citation` annotations. Count the queries in `google_search_call` to know what a prompt cost. The example is Google's, with the model id changed to Argon; we have not run it against Argon.

Google's docs also point to its terms for how the search suggestions must be shown. Read them before you hide the suggestions in your product.

What grounding costs at your volume

The free 5,000 queries cover a small agent. Past that, the price depends on how many searches each prompt triggers, which you control only through the prompt and the tools you offer.

10,000 prompts in a month, 1 search each
  10,000 queries - 5,000 free = 5,000 billed  ->  5,000 x 14 / 1,000 = 70 USD

10,000 prompts in a month, 3 searches each
  30,000 queries - 5,000 free = 25,000 billed ->  25,000 x 14 / 1,000 = 350 USD

The free allowance is shared across Gemini 3 and newer models, so a second agent on another of those models draws from the same 5,000. The Argon token bill comes on top, and that bill doubles on 2026-12-01: see Gemini 4 Argon pricing.

Grounding with Google Search alternatives

The main alternatives are the search tools built into other model APIs and a search API your agent calls as its own tool. OpenAI and Anthropic charge less per 1,000 searches but give no free allowance. An outside search API works with any model and can return data grounding does not, such as ranked positions.

A search API called as a tool is billed per call by its provider. Here is what the web search job and the Google results job cost today, read from looot's live catalog:

The prices of each search vendor's own plans, from Perplexity to Brave and Tavily, are in Web search API pricing per 1,000 searches.

When grounding is the right choice, and when it is not

Grounding is the simplest choice when you run Gemini, stay under 5,000 searches a month, and want an answer with cited Google sources. Pick an outside search tool when you pass that volume at several searches per prompt, when you need the result list itself (positions, domains, snippets) rather than a written answer, or when the same agent also runs on GPT-6.1 Sol or Claude Opus 5.5. Grounding only works inside Gemini.

Many agents use both: grounding for quick answers, a data tool for records. Gemini 4 Argon MCP shows how to add looot's server next to grounding, and the Gemini 4 Argon API guide covers the rest of the tools list.

What it cannot do

• looot does not sell Grounding with Google Search. Grounding is billed by Google; looot's search jobs are a separate tool your agent calls.
• Gemini 4 Argon is not in looot's catalog today either. You pay Google for the model and for grounding.
• A search job returns results, not a grounded answer with Google's citations. The model still writes the answer.
• It cannot lower Google's price or move your free allowance.
• It does not run on a timer. A recurring search report needs scheduled runs, not available yet.