How does Google Cloud's cost estimator agent work?

Per Google's how-to page, you open Cost estimation in the Billing section of the Google Cloud console, describe the agent in plain English and confirm one of four sample designs. You then edit assumptions such as users, sessions, turns, tool calls, models, tokens and caching share, press Generate estimate and download a CSV. The page says a pricing engine, not the AI, calculates the totals from your billing account's price sheet. We did not run the tool.

Is the estimate a price I can rely on?

Google's page calls estimates "non-binding planning projections" and says an estimate "does not lock in a fixed monthly price or cap your spend". It adds that estimates are only as accurate as the usage assumptions you enter. The pages we read give no accuracy figure and no comparison with real invoices.

What does the estimator not cover?

During Preview it prices only the Gemini Enterprise Agent Platform stack. The page says you cannot add BigQuery, Cloud SQL, Google Kubernetes Engine, Cloud Storage or Pub/Sub; for those it points to the Google Cloud pricing calculator. It works in English only, does not look up past bills, and excludes human labor, licenses and third-party tools.

What numbers decide the cost of an AI agent?

Google's page lists users, sessions, turns and tool calls, the model, input and output tokens, the share of input cached, the share of requests grounded in search, runtime hours, session storage, safety scanning, evaluation runs, search volume, and trace, log and metric volumes. In our reading, ask any vendor to quote against the same assumptions in writing.

Is the cost estimator agent ready for production budgeting?

Google labels it Preview. Its Pre-GA Offerings Terms say such offerings are provided "as is", can change without notice and are not covered by any SLA or Google indemnity, and that no data processing terms apply unless a notice or documentation says otherwise. Google's own page calls the output a planning projection. These are standard terms; a negotiated agreement may differ. We are not lawyers; ask your counsel.

Google Cloud Cost Estimator Agent: Limits, Preview

A glass-framed calculator with brushed-metal keys, a blank dark screen and one red key, with a round dial gauge behind it showing a glowing red arc and needle. Four small glass cubes, each holding a stacked metal-and-red block, feed the dial through thin red cables, and a stack of plain metal discs stands beside the calculator on a dark reflective floor.

The short version

Google Cloud's cost estimator agent, labelled Preview, turns a plain-English agent description into a downloadable monthly estimate: you confirm one of four sample designs and edit assumptions such as users, tool calls and tokens. Google's page says the figures are non-binding, do not cap spend and cover only Google's own agent platform, in English. Google's standard Pre-GA terms (rules for pre-release features) apply, including no uptime promise.

In plain words

Think of this tool as a quote form that fills itself in (our comparison, not Google's). You describe the AI agent you want to build, such as a helpdesk for 5,000 internal users; the tool suggests the closest of four sample designs, which you confirm, and fills in starting numbers, which you change. It adds up a monthly figure from your own Google Cloud price sheet. Google calls the result a non-binding planning projection, not a binding price quote, and it covers only Google's own agent platform for now. Google labels the tool Preview; its launch-stage page says Preview offerings are, unless Google says otherwise, intended for use in test environments only. We tested nothing.

Before you build an AI agent, how do you know what it will cost to run once real people use it? Google Cloud now has a tool in its Billing console, labelled Preview, that turns a plain-English description of an agent into a monthly estimate and a downloadable CSV file. Google sells the platform whose costs the tool estimates.

What numbers set an AI agent's bill?

Google's page says an agent's cost is not estimated by picking fixed server shapes or storage tiers, but from usage. It names three areas of use to account for: token usage (think of a token as a small piece of text that a model reads or writes), how the agent behaves (how often it calls tools or hands work to helper agents) and supporting infrastructure (search, runtime compute, stored sessions, safety scanning, evaluations and logs).

Here are the main inputs, as Google lists them. The first two rows are about traffic and model choice; the last three are more technical, and in our reading they are the rows to hand to whoever builds the agent.

Area	Main inputs, as Google lists them
Traffic	Users, sessions, turns (exchanges inside one conversation, our gloss) and tool calls
Model and tokens	Which Gemini model; input tokens; the percentage of input tokens cached; output tokens
Grounding	The percentage of requests grounded with Google Search, Enterprise Web and Google Maps (grounding means tying answers to a search source, our gloss)
Runtime and sessions	Runtime hours (vCPU and RAM hours), session read/write vCPU hours and session storage, and Agent Gateway requests
Safety, search and logs	The share of tokens scanned by the safety scanner (Model Armor), monthly evaluation runs and their token volume, monthly search request volume, and trace spans, log storage and monitoring metrics volume

In our reading, this list works as a checklist for any vendor's quote, not only Google's. Ask each supplier to quote against the same assumptions in writing. That makes quotes easier to compare, but not identical: also ask which services, usage units, discounts and outside costs each quote includes.

What does the tool do?

Google's release notes list it under November 10, 2026, as "Estimate costs for AI agents using the new cost estimator agent (Preview)". The how-to page prints "Last updated 2026-11-10 UTC."

The flow, as the how-to page describes it:

In the console's Billing section, open Cost estimation and describe the agent.
The tool matches your description to one of four sample designs and asks you to confirm: Conversational Chatbot, Workflow Operator, Autonomous Researcher or Multi-agent Orchestrator.
It fills a form with starting values and any numbers from your prompt. You edit the form or ask in chat.
You press Generate estimate. You get a total estimated monthly cost, a breakdown by component and a CSV file (cost_estimate.csv) with the architecture total and SKU-level details (SKU: a price-list item, our gloss) for each priced item. The page lists fields such as SKU ID, region, usage unit, list cost and estimated cost.

Who does the arithmetic?

The page says a pricing engine calculates the totals from your own Cloud Billing price sheet, including contracted prices, and that "The AI agent never guesses SKU prices or calculates totals on its own." The page says you need a role on the Cloud Billing account (Billing Account Viewer, Billing Account Administrator or a custom role). That role, it says, lets you "view all costs in the billing account". In our reading, decide who gets it before you hand it to someone who only needs an estimate.

How far can you trust the number?

Estimates "serve solely as non-binding planning projections", and the page adds: "Estimates are only as accurate as the usage assumptions entered." It also says an estimate "does not lock in a fixed monthly price or cap your spend".

The pages we read give no accuracy figure and no comparison with real invoices. The tool cannot look at your past bills or create budgets. The page says to monitor your metered spend after you deploy and to set Cloud Billing budgets and spend caps as needed. Spend caps are a separate Google feature, and Google's spend-cap page says enforcement is not instant and overage is billed as normal.

What does it leave out?

Item	What Google's how-to page says
Which services	During Preview it prices only the Gemini Enterprise Agent Platform stack. You cannot add BigQuery, Cloud SQL, Google Kubernetes Engine, Cloud Storage or Pub/Sub; the page points to its pricing calculator for those.
Language	The tool "only supports interactions in the English language".
Costs outside the tool	Human labor, licenses, process mapping, re-architecting, grounding data stores and third-party tools are not covered (the page does not define "grounding data stores").
Models	Examples: Gemini 3.1, 3.5 and 4.0, and later Flash and Pro releases. The page does not say whether Gemini 4 Hydrogen or Gemini 4 Carbon can be picked (models in our Nov 11 and Nov 12 editions).

The pages we read do not state a price or quota for the estimator itself.

What does "Preview" mean in the fine print?

A label on the how-to page says the feature is subject to the "Pre-GA Offerings Terms" in Google's Service Specific Terms, and that Pre-GA products are available "as is" and "might have limited support". Google's launch-stage page says Preview offerings are "intended for use in test environments only" unless Google says otherwise. The pages we read give no general-availability date for this tool.

Here is what section 5 of the Terms says. These are the standard terms as printed on Google's page (marked "Last modified November 11, 2026"). They refer to "the Agreement", and a customer with a negotiated agreement may have different terms. The Terms page does not name Cloud Billing; the link comes from the how-to page's label. We are not lawyers; ask your counsel.

Changes. Pre-GA Offerings "may be changed, suspended or discontinued at any time without prior notice".
Promises. They are provided "as is" and are "not covered by any SLA or Google indemnity". An SLA is a written uptime promise; an indemnity is a promise to cover certain legal claims (both glosses ours). Unless a notice or documentation says otherwise, the Terms also exclude TSS (we read this as technical support, our gloss).
Liability. Google's liability for Pre-GA Offerings is limited to the lesser of the cap in the customer's agreement or $25,000. Four matters are listed where the cap does not affect any exclusions the customer's agreement already has, among them death or personal injury from negligence, infringement of intellectual property rights and "fraud or fraudulent misrepresentation".
Data. Unless a written notice or Google documentation says otherwise, "no data processing terms (including the Cloud Data Processing Addendum) apply to Pre-GA Offerings". Google adds that customers "should not use Pre-GA Offerings to process personal data" or other regulated data. The how-to page we read indicates no exception.

What happens to the prompts you type?

The pages we read do not state what Google keeps from the prompts you type into the estimator. We searched two pages for the estimator and found no mention: the Google Cloud Privacy Notice and the page on how Gemini products in Google Cloud use your data. That Gemini page says that when you use Gemini in the Google Cloud console, Google Cloud handles your prompts under its terms of service and the Cloud Data Processing Addendum, and that Gemini does not use your prompts or its responses to train its models. The Terms above say no data processing terms apply to Pre-GA Offerings unless a notice or documentation says otherwise. The pages we read do not say which applies to prompts typed into the estimator, or where they are processed, so we draw no conclusion. In our reading, type workload numbers into it, not personal or regulated data.

What is unclear about price changes?

The pages we read do not say how an estimate treats a price that changes on a set date, or whether the CSV records the date a price applies (its listed fields include none). Our Nov 11 edition covers a Gemini price step dated December 1.

What to do this week

Four checks for an owner or team lead. They are our suggestions, not Google's advice, and not legal or buying advice.

Write down the numbers on the list above for the agent you have in mind: users, sessions, tool calls, model, tokens, how much is cached. Ask any vendor to quote against the same figures.
If you run on Google Cloud, ask your Billing Account Administrator who holds the role the estimator needs, since the page says it can view all costs in the billing account.
Treat the output as a planning figure. After launch, compare it with your first month's metered bill (our suggestion). Google's page says to monitor metered spend and set budgets and spend caps as needed.
In our reading, keep personal or regulated data out of the prompt, and ask your counsel what the Pre-GA Offerings Terms mean for your own agreement.

Book an AI discovery call

How we know

On November 12, 2026 we opened the seven pages under Sources. We searched two of them (the Privacy Notice and the Gemini data page) for the estimator, read the Gemini page's passage on prompts, and did not read the Privacy Notice end to end. We did not open the console, run the tool, read the agent platform's pricing pages (so we cite no Google price) or read the customer Agreement the Terms refer to. Pages can change.

Google Cloud's Cost Estimator Agent (Preview): How It Estimates an AI Agent's Monthly Cost, What It Leaves Out, and What the Terms Say