Research-backed calculator guide
Vertex AI Grounding Cost per Query: Search, Maps, Data and Token Charges
Calculate Vertex AI grounding cost per request by separating Search or Maps query fan-out, free allowances, grounding meters, and Gemini token charges.
Quick answer
Vertex AI grounding cost per query is not one universal surcharge. Price a request as separate meters: Gemini input and output tokens, plus the grounding activity that actually happened. For current Gemini 3 and newer Search grounding, Google documents 5,000 free Search queries per month shared across eligible Gemini 3/4 models, then $14 per 1,000 Search queries. Maps uses the same 5,000-query monthly allowance and $14 per 1,000 queries in the current Agent Platform table. Grounding with your data is listed separately at $2.50 per 1,000 prompts. These are USD list-price observations checked on 2026-11-02, not a quote for your contract or region. [1] [2]
The word *query* is the important part. One end-user prompt can cause Gemini to issue zero, one, or several Search queries. Google’s documentation says the current Search model can generate multiple queries, and current pricing charges each individual query. That means “10,000 grounded prompts” and “10,000 grounding queries” are different inputs to a budget. [5]
A practical monthly estimate is:
monthly cost = model input tokens + model output tokens + billable Search queries + billable Maps queries + grounding-with-your-data prompts + any other enabled service meters
For an answer that uses Search grounding, count the user prompt’s model tokens and the model’s response tokens separately from the retrieved Search material. Google’s current pricing page says input tokens provided by Search or Web Grounding are not charged as model input; your own prompt and the generated answer still need to be measured under the selected model’s token rates. [1]
The four questions that prevent a bad estimate
1. Which grounding path is enabled?
“Grounding” is an umbrella term rather than a SKU. Google describes Search, Maps, Agent Search, RAG Engine, Elasticsearch, a customer search API, and other paths as different ways to connect model output to verifiable information. Search is a web-index path. Maps is location and place data. Agent Search connects to website, structured, or unstructured data. RAG Engine is another managed retrieval path. [3]
Do not combine these meters merely because they all improve factuality. First record the tool or service in the request log. A request that calls Google Search and then your own retrieval service can incur both service costs, while a request that only uses a model with a grounding tool may not use Agent Search at all.
2. Are you budgeting prompts or underlying queries?
For Gemini 3 and newer Search grounding, the current billing unit is the individual Google Search query. Google gives a simple operational sequence: the model analyzes the prompt, generates one or multiple queries when useful, processes the results, and returns a cited response. The API’s response can expose the generated query list in grounding metadata. [5]
Maps behaves similarly. Google’s Maps documentation gives examples where one prompt creates three place-search queries, or one place query plus three directions queries. Each is a Google Maps query for billing purposes. [6] A per-prompt spreadsheet that assumes one query per request will understate cost when fan-out occurs.
Older Gemini 2.5 and earlier Search rows use a different commercial unit: Google’s Gemini API documentation says those models are billed per prompt, even if multiple Search queries are sent. The model generation and the account’s pricing table therefore belong in the input data for the calculator. [5]
3. Has the account-level free allowance been consumed?
The current Gemini 3/4 Agent Platform pricing table lists 5,000 free Search queries per month shared across those model families and 5,000 free Maps queries per month. After the allowance, the listed price is $14 per 1,000 queries. The allowance is not a free 5,000 queries for every prompt, model, API key, or region; treat it as an account-level monthly pool unless your contract or billing export says otherwise. [1]
Grounding with your data is a different row: $2.50 per 1,000 prompts. Agent Search has its own pricing model. Its general pricing page describes 10,000 queries per account per month at no cost for the free trial, then lists Standard Search at $1.50 per 1,000 queries and Enterprise Search at $4 per 1,000 queries, with advanced generative answers as an add-on. Those are Agent Search meters, not interchangeable Search-grounding prices. [4]
4. Which model token rates and date apply?
Grounding fees do not replace inference fees. Every request still needs a model, token counts, and a rate card. The same Google pricing page contains model-specific rates and time-bound introductory prices. For example, its current table lists Gemini 3.8 Flash introductory input/output rates of $0.75/$3.75 per 1 million tokens through December 31, 2026, with standard rates beginning January 1, 2027. [1]
Record the model version, endpoint or product, region, traffic tier, batch or priority mode, and the rate-card date in your ledger. If you use Gemini API rather than Vertex AI or Agent Platform, use the Gemini API pricing page for that product because Google explicitly notes that prices may differ between the Gemini API and Gemini Enterprise Agent Platform. [2]
A request-level cost formula
Start with token costs:
input_token_cost = input_tokens / 1,000,000 × input_rate
output_token_cost = output_tokens / 1,000,000 × output_rate
Then calculate grounding after allocating the monthly allowance:
billable Search queries = max(0, Search queries this month − unused Search allowance)
Search grounding cost = billable Search queries / 1,000 × Search rate
Use the same pattern for Maps. For grounding with your data, use prompts rather than underlying retrieval hits when the applicable pricing row is prompt-based:
your-data grounding cost = billable prompts / 1,000 × $2.50
Finally, add any Agent Search query, storage, index, vector, logging, or application infrastructure meters that your architecture actually enables. Do not hide them inside “grounding cost”; separate lines make reconciliation possible.
The formula is an estimate until you replace assumptions with exported usage. A calculator should accept observed average Search queries per prompt and a percentile or distribution, not just a binary “grounding on” switch. For operations planning, store at least prompt ID, model, grounding type, generated-query count, token counts, response status, and project or cost center.
Worked cost example: 10,000 Search-grounded requests
Assume a US-dollar, on-demand estimate checked 2026-11-02. The workload uses Gemini 3.8 Flash on the current introductory rate shown in the Agent Platform pricing table: $0.75 per 1 million input tokens and $3.75 per 1 million output tokens. Each end-user request has 1,000 billable model input tokens and 500 output tokens. The model generates an average of 1.2 Google Search queries per request. Assume the shared monthly Search allowance has not been used before this workload.
| Step | Assumption or input | Calculation | Output |
|---|---|---|---|
| 1. End-user prompts | 10,000 prompts | Given | 10,000 prompts |
| 2. Search fan-out | 1.2 queries per prompt | 10,000 × 1.2 | 12,000 Search queries |
| 3. Free Search allowance | 5,000 queries | 12,000 − 5,000 | 7,000 billable queries |
| 4. Search grounding | $14 per 1,000 queries | 7,000 / 1,000 × $14 | $98.00 |
| 5. Model input | 10,000 × 1,000 tokens | 10,000,000 / 1M × $0.75 | $7.50 |
| 6. Model output | 10,000 × 500 tokens | 5,000,000 / 1M × $3.75 | $18.75 |
| 7. Estimated total | Search plus model tokens | $98 + $7.50 + $18.75 | $124.25 |
| 8. Effective request cost | Total / 10,000 prompts | $124.25 / 10,000 | $0.012425 |
The result is about 1.24 cents per end-user prompt under these assumptions. The Search line is $98 because the workload creates 12,000 underlying queries, not because there are 12,000 end-user prompts. If the same 10,000 prompts generated one query each, only 5,000 would be billable after the allowance and Search cost would be $70, reducing the estimated total to $96.25. If the account had already consumed the allowance, all 12,000 queries would be billable and Search cost would be $168, taking the total to $194.25.
This example intentionally excludes taxes, negotiated discounts, committed-use arrangements, retries, failed requests, storage, Agent Search, embeddings, logging, network egress, and application compute. It also assumes the cited Gemini 3.8 Flash introductory rates remain applicable on the checked date. Replace the token rates after a scheduled price change; do not carry an introductory rate into 2027 without rechecking the official table.
Search, Maps, and your data: compare the meters
Google Search grounding
Search is best modeled from actual grounding metadata. Log the generated query array or the provider usage field, then calculate the monthly distribution. If your application uses dynamic retrieval, Google’s Gemini API documentation says only requests that contain at least one grounding support URL are charged for Search grounding in that API context; model token costs still apply. [2] That qualification makes “tool enabled” an unreliable proxy for “billable query.” Verify the product’s billing export and terms for your endpoint.
Google Maps grounding
Maps cost is driven by place, routing, or search-along-route operations that the prompt triggers. A request asking for restaurants may fan out across several place searches. A distance question can include place resolution and multiple directions queries. Capture the count by request and preserve latitude, longitude, language, and grounding type when relevant, because these details affect behavior and make a later audit possible. [6]
Grounding with your data and Agent Search
For managed grounding with your data, the current Agent Platform row is $2.50 per 1,000 prompts. That is not the same as Agent Search’s Standard or Enterprise query price. Agent Search can also charge for advanced generative answers and indexed data. Its pricing page defines a query as an API call, directly or through an integration or console, and illustrates that a Search Enterprise query plus an advanced-answer add-on produces separate query lines. [4]
For a private corpus, compare the full architecture, not just the retrieval line: indexing or storage, embedding generation, retrieval, model input, output, cache behavior, and refresh frequency. The RAG, embedding and vector database cost calculator is a useful companion for those non-grounding retrieval components.
How to reconcile estimates to an invoice
Build a ledger with one row per model request and separate columns for prompt_id, model, grounding_type, end_user_prompts, generated_search_queries, maps_queries, your_data_prompts, input_tokens, output_tokens, free_allowance_applied, and rate_card_date. Aggregate by project, feature, customer, and day. Compare the resulting totals with the billing export by SKU rather than expecting a single “grounding” SKU to explain every line.
Investigate three variances first. A higher-than-expected Search bill usually means query fan-out or an already-consumed allowance. A higher token bill can mean longer context, larger retrieved instructions, thinking tokens, or output growth. A higher Agent Search bill can reflect advanced answers, index storage, or a configurable capacity commitment rather than ordinary model grounding.
For a broader model-rate comparison before selecting a Gemini tier, see Model pricing comparison. If you are comparing hosted grounding with self-hosted retrieval or inference, the Self-hosted GPU inference cost calculator can keep the infrastructure alternative separate from Google’s managed-service meters. For a product-level view that includes workflow labor and build scope, use the AI agent development cost estimator.
FAQ
Is Vertex AI grounding charged per prompt or per query?
It depends on the model generation and product path. Current Gemini 3 and newer Search and Maps rows charge per underlying Search or Maps query, so one prompt can create several billable units. Google documents Gemini 2.5 and older Search grounding as prompt-based in the Gemini API documentation. Grounding with your data is listed as a per-prompt price. Confirm the exact model and endpoint before applying a formula. [1] [5]
Does one user prompt always create one Google Search query?
No. The model may decide that no search is useful, or it may generate multiple queries. Google’s Search documentation describes automatic query generation and exposes the query steps in the response. Measure observed query counts rather than multiplying prompt volume by one by default. [5]
Are retrieved Search tokens billed as ordinary input tokens?
Google’s Agent Platform pricing table says input tokens provided by Google Search or Web Grounding are not charged. That does not make the whole request free: your prompt and the model’s output remain subject to the selected model’s rates, and other tools may have their own billing rules. Treat the retrieved-token exclusion as a product-specific pricing rule, not a general token-billing assumption. [1]
What is the current free allowance for Gemini 3 Search grounding?
The current Agent Platform table lists 5,000 Grounding Queries per month at no charge, aggregated across Gemini 3 and Gemini 4 models, with $14 per 1,000 queries above the allowance. The Gemini API page also lists 5,000 free Search requests per month for Gemini 3 and newer. The product, account, model, and rate-card date matter, so recheck the official page before publishing a forecast. [1] [2]
How should I estimate Google Maps grounding cost?
Count individual Maps operations, not merely user prompts. A single request can produce multiple place searches or a place lookup followed by several directions queries. Apply the current allowance and per-query rate to the monthly Maps total, then add Gemini token cost. Log the generated Maps query count and grounding type so the forecast can be reconciled. [6]
Is grounding with your data the same as Agent Search pricing?
No. Grounding with your data and Agent Search are distinct pricing rows and product paths. The current Agent Platform table lists grounding with your data at $2.50 per 1,000 prompts, while Agent Search lists separate Standard, Enterprise, advanced-answer, capacity, and storage charges. Compare the architecture and meters that your request actually uses. [1] [4]
Run your scenario
Turn this framework into an editable estimate.
Use dated, source-linked assumptions and see cost per request, successful task and month.
Limitations and assumptions
This guide is a calculation method, not a quote. Prices and allowances are time-sensitive and can vary by product, model generation, endpoint, region, tier, contract, and discount. The figures above are USD list-price observations checked on 2026-11-02. They exclude taxes, negotiated pricing, retries, failed calls, concurrency capacity, quota changes, storage, indexing, embeddings, logging, network transfer, application compute, and downstream business costs. A real budget should use exported usage and the billing SKU definitions for the project.
The Search benchmark for this article used the exact query “Vertex AI grounding cost per query pricing” in English, United States on 2026-11-02. The search interface did not reliably expose sponsored labels, so the benchmark records that limitation and does not make a universal organic-ranking claim. Results included official documentation, independent pricing explainers, and community discussions. The observed opportunity is the request-level ledger and worked fan-out calculation above; there is no claim that the topic is easy to rank or that traffic will follow.
Primary sources checked
Pricing and product behavior can change. These sources support the time-sensitive claims in this guide; verify current provider terms before committing spend.