Vertex AI Pricing: Gemini Rates and Two Monthly Budgets

Vertex AI pricing puts a 100,000-call support-drafting workload at an effective $337.50 a month on Gemini 3.8 Flash today, or $675 after its introductory promotion ends. Google now calls the platform Gemini Enterprise Agent Platform. Start with pay-as-you-go, count reasoning as output, and budget against the post-promotion price before committing to production.

Verified against Google Cloud's and the Gemini Developer API's live pricing pages, October 2026, on 20 October. These are USD public rates and calculated examples, not a production benchmark or a negotiated quote. Google's Cloud promotion uses credits back on eligible net spend, so the effective cost below is not a promise that your initial invoice will be that amount. Google Cloud pricing

Vertex AI Pricing at a Glance

Vertex AI charges for the model, the tokens and the serving option you choose. A token is a small unit of text or encoded media. Input includes what you send; output includes the response and billable reasoning. A cheap input rate can still produce an expensive answer if the model generates substantial reasoning.

The table shows Standard inference, USD per million tokens, using Global rates where the page specifies a region. Input here means text; modality differences follow below. For Pro rows, values separated by a slash mean input context up to 200,000 tokens / above 200,000 tokens. The threshold applies to each request, not monthly volume. Gemini 2.5 rows are presented without a Global label because their source table does not specify one. Official model tables

The missing Gemini 3 Flash Preview output cell is deliberate. Google's current Cloud Standard table omits it, although its Priority and Flex/Batch tables list output rates. Importing a price from a different API would conceal that gap. Confirm that Standard SKU before budgeting it.

For a new text feature, evaluate a Flash-Lite model first if it meets your quality bar. Move to a more expensive model when it measurably improves accepted answers, reduces retries or handles a task the cheaper model cannot. Token rates alone establish the cost floor; they do not establish answer quality.

What the Rename Changes for Your Budget

The developer platform formerly called Vertex AI is now Gemini Enterprise Agent Platform. On April 22, 2026, Google wrote: “Gemini Enterprise Agent Platform is our new developer platform and evolution of Vertex AI.” Its release notes record the naming changes on the same date. Google's announcement, release notes

Google distinguishes the Agent Platform, where technical teams build and operate agents, from the Gemini Enterprise app, the employee-facing application. A seat subscription and a model's token bill answer different purchasing questions. The rename itself is not a reason to substitute an app subscription price for an API estimate.

Vertex AI Studio Pricing

Vertex AI Studio became Agent Studio in Google's naming changes. When estimating Gemini model calls from that environment, use the model's Cloud inference rates and any separately used services. Do not treat the interface's old name as a separate unlimited-token plan. For personal Gemini app subscriptions, use the site's Gemini pricing guide. Naming changes

Vertex AI Model Pricing Beyond the Headline Rate

Choose the endpoint and modality before applying a rate. “Gemini input” can mean text, audio or images, and some models have different meters for each. The prices in this section come from the same Google Cloud pricing page.

Global and Non-Global Endpoints

The page lists a higher non-global rate for several current models. For Standard input/output, per million tokens:

• Gemini 4 Argon: $2.20/$11 through November 30, then $4.40/$22. Non-global cached input is $0.11, then $0.22. Argon Cyber Permissive uses $4.40/$22 and $0.22 cached input without the Argon promotion.
• Gemini 3.8 Flash, 3.7 Flash and 3.6 Flash: $0.825/$4.125 through December 31, then $1.65/$8.25. Cached input is $0.0825, then $0.165. 3.8 Flash Cyber uses $1.65/$8.25 and $0.165 cached input.
• Gemini 3.5 Flash: $1.65/$9.90, with $0.165 cached input. 3.5 Flash-Lite: $0.33/$2.75, with $0.033 cached input.
• Gemini 3.1 Flash-Lite: $0.275/$1.65 for text/image/video input and text output, with $0.0275 cached input. Audio input is $0.55, or $0.055 cached.

These listed non-global rates are 10% above their corresponding Global rows. That is a useful budgeting rule for these rows, not permission to invent a regional price for a model that lists only Global. Choose the endpoint that meets your deployment requirements, then price that endpoint.

Audio Input, Live Sessions and Speech Models

For Global Standard inference, Gemini 3.1 Flash-Lite audio input is $0.50, or $0.05 cached, per million tokens. Gemini 3 Flash Preview audio input is $1.00, or $0.10 cached. Gemini 3.5 Flash, 3.5 Flash-Lite, 3.6/3.7/3.8 Flash, 3.8 Flash Cyber, 3.1 Pro Preview and Argon use their listed input rates across the modalities named in their rows.

Gemini 2.5 Flash Lite audio input is $0.30, or $0.03 cached. The Cloud table displays Gemini 2.5 Flash audio input as $1.00 at up to 200K context and $0.30 above 200K, with $0.10 cached in both columns. That unusual Standard row is reproduced as published; get confirmation before building a long-audio forecast around the lower value.

Gemini 3.8 Live API, Google's live interaction model, is listed for Non-global use at $0.75 text input, $1 image/video input, $3 audio input, $4.50 text output, $12 audio output and $1 avatar-video output, all per million tokens. Its cache columns are N/A. The page says accumulated session context is processed and charged again on each turn. A ten-turn conversation therefore cannot be priced as if the original context were sent only once.

Gemini 2.5 Flash Live API lists $0.50 text input, $3 image/video or audio input, $2 text output and $12 audio output per million tokens. Its Priority and Flex/Batch cells are N/A.

For speech-specific jobs, Gemini 3.8 Flash TTS Preview lists $0.50 text input/$9 audio output, and Gemini 3.8 Flash-Lite TTS Preview lists $0.50/$6, in the page's per-million-count units through December 31, 2026. From January 1, 2027 they become $1/$18 and $1/$12 respectively. The page describes the introductory reduction as promotional credits.

Gemini 3.5 Live Translate lists $3.50 audio input/$21 audio output. Gemini 3.5 Transcribe Live lists $3.50 audio input/$21 text output; Gemini 3.5 Transcribe lists $2/$12. Those are per-million-token rates, not a flat charge per conversation. Use the model's own audio conversion assumptions before turning them into minutes.

Image, Video and Embedding Models

Image generation has a separate output meter. The Cloud Standard rates below are input / text output / image output, in USD per million tokens:

• Gemini Nano Banana 2.1: $1.50 / $7.50 / $30, with $0.15 cached input.
• Gemini 3.1 Flash-Lite Image, Nano Banana 2 Lite: $0.25 / $1.50 / $30, with $0.025 cached input.
• Gemini 3.1 Flash Image, Nano Banana 2: Global $0.50 / $3 / $60, cached $0.05; Non-global $0.55 / $3.30 / $66, cached $0.055.
• Gemini 3 Pro Image, Nano Banana Pro: $2 / $12 / $120, with $0.20 cached input.
• Gemini 2.5 Flash Image: $0.30 / $2.50 / $30; cached input is N/A.

Resolution changes the number of image tokens. For example, Google lists approximate Nano Banana 2.1 output costs of $0.034 at 1K, $0.05 at 2K and $0.113 at 4K. Prompt input remains a separate charge. An image output rate should never be applied to text output or interpreted as a per-image price.

Gemini Omni Flash and Gemini Omni 1.1 Flash list $1.50 input, $9 text output and $17.50 video output per million tokens. The page's video-token conversion varies by resolution, so a fixed token rate does not imply a fixed price per second across resolutions.

Gemini Embedding 2, Unified Multimodal Preview, which converts content into numerical representations for retrieval, lists online input rates of $0.20 text, $0.45 image, $12 video and $6.50 audio per million tokens. Batch input rates are $0.10, $0.225, $6 and $3.25 respectively; output tokens have no charge. The separate Gemini Embedding row is printed as $0.00015 online or $0.00012 Batch per 1,000 count, with no output charge. Keep that source unit attached to the rate when verifying the SKU.

Older Price Rows Are Not a Deployment Guarantee

Google still publishes Gemini 2.0 pricing: Flash text input/output at $0.15/$0.60, audio input $1, and Batch at $0.075/$0.30, audio input $0.50. Flash Lite is $0.075 input/$0.30 output, including its listed audio input, with Batch $0.0375/$0.15. All are per million tokens.

Its 2.0 Flash Image Generation row lists $0.15 input, $1 audio input, $0.60 text output and $30 image output; its 2.0 Flash Live API row lists $0.50 text input, $3 audio/video/image input, $2 text output and $12 audio output. Blank Batch cells are not advertised discounts. The page also retains older Gemini 1.5 and 1.0 modality schedules. A retained pricing row alone does not establish that a model is available for your new deployment; verify the model ID before selecting it.

Batch, Priority and Caching: Which Discount Applies?

Use Batch for work that can wait, and treat caching as a separate calculation. Batch submits work for asynchronous processing. Caching reuses prior input context. A lower price in one column does not establish that both discounts combine on every model.

The Listed Batch and Flex Rates

Google labels Gemini 4's discounted table Batch, and the relevant Gemini 3 and 2.5 tables Flex/Batch. Listed Global input/output rates per million tokens are:

• Gemini 4 Argon: $1/$5, cached $0.05, through November 30; then $2/$10, cached $0.10. Argon Cyber Permissive: $2/$10, cached $0.10.
• Gemini 3.8 Flash, 3.7 Flash and 3.6 Flash: $0.375/$1.875, cached $0.0375, through December 31; then $0.75/$3.75, cached $0.075. 3.8 Flash Cyber: $0.75/$3.75, cached $0.075.
• Gemini 3.5 Flash: $0.75/$4.50, cached $0.075. 3.5 Flash-Lite: $0.15/$1.25, cached $0.015.
• Gemini 3.1 Flash-Lite: $0.125/$0.75 for text/image/video input and text output, cached input $0.0125; audio input $0.25, cached $0.025.
• Gemini 3.1 Pro Preview: $1/$6 up to 200K context, $2/$9 above it. Its Cloud Flex/Batch cached-input cells are N/A.
• Gemini 3 Flash Preview: $0.25 text/image/video input, $0.50 audio input and $1.50 text output; cache cells are N/A.
• Gemini 2.5 Pro: $0.625/$5, or $1.25/$7.50 for long context. 2.5 Flash: $0.15/$1.25, with $0.50 audio input. 2.5 Flash Lite: $0.05/$0.20, and its Cloud table explicitly lists $0.05 audio input.

The corresponding listed Non-global discounted rates for Argon, the newer Flash models and 3.1 Flash-Lite are 10% above the Global figures. For example, current 3.8 Flash is $0.4125/$2.0625 with $0.04125 cached input, while 3.1 Flash-Lite is $0.1375/$0.825 with $0.01375 cached text input. Google's tables, including N/A cells, remain the authority. Batch and Flex pricing

For image models, the discounted input/text/image-output triples are Nano Banana 2.1 $0.75/$3.75/$15, Nano Banana 2 Lite $0.125/$0.75/$15, Nano Banana 2 $0.25/$1.50/$30, Nano Banana Pro $1/$6/$60, and 2.5 Flash Image $0.15/$1.25/$15. The first four have cached-input prices of $0.075, $0.0125, $0.025 and $0.10. Google's Nano Banana table groups Flex/Batch/Off-peak but explicitly says Nano Banana 2.1 does not support Flex.

Priority Is a Premium

For the text-model rows that have Priority prices, the listed input, output and cached-input rates are 1.8 times Standard. For example, current Global 3.8 Flash is $1.35 input/$6.75 output/$0.135 cached input, and 3.1 Pro Preview is $3.60/$21.60/$0.36 for short context or $7.20/$32.40/$0.72 for long context. Argon's current Global Priority rates are $3.60/$18/$0.18, becoming $7.20/$36/$0.36 after its promotion.

The four Nano Banana model rows also list Priority at 1.8 times their Standard rates, including image output. Gemini 2.5 Flash Image, 2.5 Flash Live API and 2.5 Pro Computer Use-Preview show N/A Priority cells on this page. Do not apply a multiplier to a service option the table does not list.

Caching Has a Storage Bill

Cloud's explicit-cache storage table lists $0.000001 per token-hour, equivalent to $1 per million token-hours, for its listed Flash and Flash-Lite models. It lists $0.0000045 per token-hour, equivalent to $4.50 per million token-hours, for Gemini 3.1 Pro, 3 Pro and 2.5 Pro. Cached-input processing is charged separately. The storage table does not list an Argon explicit-cache rate; do not manufacture one from another model. Cache storage prices

Consider a reusable 100,000-token reference pack on Global Gemini 3.8 Flash at today's rates. Processing it normally costs $0.075 per call; a cache read costs $0.0075, saving $0.0675 per hit. Retaining that pack for an hour costs $0.10, so two cache hits within that hour cover the storage charge, before any additional creation cost. That is a marginal reuse calculation, not a claim that cache creation is free.

Leave the same pack stored for an assumed 720-hour month, and storage becomes $72. It then needs 1,067 cache hits to exceed that storage bill at the same rates. Retention time matters as much as the discount percentage. Cache material your application will reuse soon, and set its lifetime accordingly.

Provisioned Throughput: Buy Capacity Only When You Need It

Provisioned Throughput reserves capacity; it is not an unlimited-token subscription. Google sells it in generative AI scale units, or GSUs. The throughput each GSU supplies depends on the model and its token mix. A monthly request total alone cannot tell you how many units to reserve. Official Provisioned Throughput pricing

The page's hourly view lists these Global / Non-global USD prices per GSU-hour:

• One-week commitment: $7.14 / $7.854.
• One-month commitment: $3.698630137 / $4.068493151.
• Three-month commitment: $3.287671233 / $3.616438356.
• One-year commitment: $2.739726027 / $3.01369863.

Google's accompanying example uses $1,200 per GSU-week, or $2,700, $2,400 and $2,000 per GSU-month for the one-month, three-month and annual terms respectively. Use the published term and billing unit; do not treat the annual rate as a cancel-any-month plan.

One Global GSU at the annual monthly rate represents $24,000 across twelve months, compared with $32,400 if the one-month rate were maintained for twelve months. That $8,400 difference only helps if you need the reserved capacity throughout the commitment. It is not a break-even against pay-as-you-go tokens because the usable GSU throughput has not yet been established for your workload.

Google also lists 50% promotional credits on eligible net Provisioned Throughput spend: Argon from October 13 through November 30, 2026, excluding Cyber Permissive; and 3.8/3.7/3.6 Flash from August 13 through December 31, 2026. The page says those credits are issued on the seventh day of the following month and expire after 30 days. They can apply to the named on-demand model SKUs and Provisioned Throughput SKUs across models; they do not change your capacity configuration or commitment burndown rates.

For intermittent traffic, start with pay-as-you-go. Revisit reserved capacity when you have measured peak demand, sustained utilization and the business cost of insufficient throughput.

Vertex AI Pricing Calculator: Two Monthly Jobs

Multiply actual billable tokens, then add the services around them. For uncached text inference:

Monthly token cost = input tokens ÷ 1,000,000 × input rate + billable output tokens ÷ 1,000,000 × output rate.

Both examples below are hypothetical workloads. Output counts include reasoning. They exclude caching, grounding, agent runtime, storage, networking, taxes and negotiated discounts. Each model call stays below the relevant long-context threshold.

Job One: Support Drafting at 100,000 Calls a Month

Assume 2,000 input tokens and 500 billable output tokens per call. Monthly volume is 200 million input and 50 million output tokens.

With Global Standard Gemini 3.8 Flash, today's promotional effective cost is:

200 × $0.75 + 50 × $3.75 = $150 + $187.50 = $337.50 a month.

That is $0.003375 per draft, or $3.375 per 1,000 drafts. At the listed rates from January 1, 2027, the same workload becomes 200 × $1.50 + 50 × $7.50 = $675. The current Non-global version is $371.25, rising to $742.50. These calculations use Google's listed model rates and assume eligible promotional credits are usable. Price inputs

A cheaper candidate changes the floor: Gemini 3.1 Flash-Lite would cost $125 for those same token counts, and 3.5 Flash-Lite would cost $185. That does not establish that either produces acceptable drafts. Evaluate the quality on your own support cases before paying for the larger model or choosing the cheaper one.

Job Two: Overnight Report Extraction

Assume 10,000 reports a month, each using 20,000 input tokens and 2,000 billable output tokens. The total is 200 million input and 20 million output tokens.

Using Global Gemini 3.1 Pro Preview below 200K context, Standard costs 200 × $2 + 20 × $12 = $640. Its listed Flex/Batch rates produce 200 × $1 + 20 × $6 = $320. That is $0.032 per report, or $32 per 1,000 reports. For an overnight job, Batch is the first serving option to evaluate. Price inputs

Together, the Standard support job and Batch report job total $657.50 in effective monthly token costs today, or $995 after the Flash promotion, holding everything else constant. Treat that as the inference subtotal, then add the rest of your application bill.

Vertex AI vs Gemini Developer API Pricing

The Gemini Developer API is not automatically cheaper for the same model. Google's developer-facing API and its Cloud platform currently list the same headline Global Standard rates for Argon, 3.8 Flash, 3.1 Flash-Lite and 3.1 Pro Preview. The model, serving tier, context length and date must all match before comparing. Developer API pricing, Cloud pricing

The differences are in the rows around that headline:

• Free access: the Developer API lists free input/output for selected models, including 3.8 Flash, but Argon and 3.1 Pro Preview are not available on its Free tier. Google's Cloud price table is a separate schedule.
• Data use: the Developer API describes Free-tier content as used to improve products and ordinary Paid-tier content as not used for that purpose. It also lists an invitation-only Contributor preview with different terms; do not mistake that for ordinary Paid pricing.
• Explicit-cache storage: the Developer API lists $0.50 per million token-hours for 3.8 Flash through December 31, then $1. Cloud's storage table lists $1. Keep each platform's storage price in its own estimate.
• Cached Batch input: Cloud's 3.1 Pro Preview Flex/Batch cells are N/A; the Developer API's Batch and Flex tables list $0.20/$0.40 for short/long context. For 3.5 Flash-Lite, Cloud lists $0.015 cached Flex/Batch input while the Developer API displays $0.02.
• Audio: Cloud's 2.5 Flash Lite Flex/Batch table lists $0.05 audio input per million tokens; the Developer API lists $0.15. Those pages disagree on that cell. Confirm the billed SKU for an audio-heavy job instead of assuming one page corrects the other.

Google positions the enterprise route with optional support, security and compliance features, Provisioned Throughput and volume discounts. Choose it when those operating requirements matter. If your application only needs model calls, compare the Developer API's simpler purchasing route on the exact features you use.

Gemini API Pricing Calculator: Compare the Same Workload

With the support job's 200M input/50M output and no caching, the Developer API's current 3.8 Flash rates also calculate to $337.50, rising to $675 on the announced date. That numerical match does not erase the Cloud promotion's credits-back mechanism or the storage differences above.

Amazon Bedrock, Amazon Web Services' managed model platform, is another deployment option. Its official Standard table lists DeepSeek v3.2 at $0.62 input/$1.85 output per million tokens in US East, N. Virginia and Ohio, and US West, Oregon. At the same assumed token counts, that is $216.50 a month. This is a price scenario for a different model, not evidence of equal answer quality. See the Amazon Bedrock pricing guide for the services that surround inference. AWS pricing

Microsoft Azure, Microsoft's cloud platform, publishes prices by model and deployment type. Select the deployment and currency on its live page before comparing it with a Global Google Cloud estimate. A price for a different context band or deployment is not a valid substitute. Azure pricing

For a higher-cost Google candidate, Argon Global Standard calculates to $900 for the support workload at its current promotional rates, then $1,800 after November 30. Compare accepted results and operating requirements before treating that as a worthwhile upgrade. Argon price inputs

Use the cheapest LLM API comparison to widen the shortlist. The decision rule is cost per accepted result: include rejected answers, repeat calls and engineering effort before declaring a provider cheaper.

Vertex API Pricing: The Costs Outside Tokens

A token estimate becomes an application budget only after you add retrieval, tools and runtime. The named billing risks here are more useful than a generic contingency percentage.

Reasoning and repeated context. Google's output rows include reasoning. Long chat histories, retrieved documents and repeated model calls increase input. For Live sessions, the page explicitly says accumulated context is charged again on subsequent turns. Count calls and their full billable context, not just users.

The long-context step. Above 200K input tokens, models with a long-context tier charge all input and output at that tier. On 3.1 Pro Preview Standard, a request with 200,000 input and 2,000 output tokens costs $0.424. At 200,001 input tokens, the same output costs $0.836004. The extra token changes the tier for the whole request. That is a concrete reason to trim retrieved material near the threshold. Context rule and rates

Grounding, or fetching supporting information. For Gemini 3 and 4, Google Search and Web Grounding for Enterprise include 5,000 grounding queries per month shared across those model families, then $14 per 1,000 queries. A single user prompt may trigger multiple queries. Google Maps has its own listed 5,000-query monthly allowance, then $14 per 1,000 queries. Grounding with your data is $2.50 per 1,000 prompts. Do not transplant the older Gemini 2.5 grounded-prompt meter into a Gemini 3/4 budget.

Failed requests and a forthcoming exception. The page generally says input/output charges apply only to HTTP 200 responses. It separately states that, from November 13, 2026, Gemini 4 completions exceeding 32K tokens will be charged for retrievable token usage even if the request ultimately fails because of a client or server timeout. A retry policy can therefore have a different cost profile for those long completions.

Tuning. Training is a separate meter, and Google says tuned inference from Gemini 3 onward costs 1.5 times the base model's prediction price. Do not use an untuned token estimate as the final budget for a tuned endpoint. Additional charges and conditions

Vertex AI Agent Builder Pricing

An agent can consume model tokens and separately billed platform services. The generative-AI page explicitly sends readers to the broader Agent Platform schedule for ML platform and MLOps charges. Price the components your agent actually uses rather than calling the token subtotal “Agent Builder cost.” Broader platform pricing

The generative-AI page does include a specific Gemini Deep Research Agent product: Standard $2 input/$12 output/$0.20 cached input, and Deferred $1/$6/$0.10, per million tokens. Those are that named agent's rates. They are not a flat fee for every agent you build, and grounding remains a separate listed meter.

Is Vertex AI Free, and Who Can Stay on Free?

Do not budget production Gemini inference on Google Cloud as an unlimited free service. Google's new-customer offer is $300 in trial credit, and it says you are not charged until activating a full paid account. Eligibility and the trial's terms still apply. A temporary credit offsets usage; it does not make the underlying token price zero. Google Cloud free offer

The Gemini Developer API separately offers free input/output on selected model rows, with limited access and model-specific restrictions. A developer experimenting within those limits, using an eligible model and accepting the Free tier's data-use terms, may never need a paid API account for that small project. A production team needing a paid-only model, higher limits or enterprise options should budget for payment. Free and Paid tiers

What to Do Before You Ship

Make the first budget from representative calls, then decide which optimizations earn their complexity. Next week, use a sample of your own workload to record input, cached input and total billed output, including reasoning. Keep the exact model ID, endpoint and serving option with each record.

1. 1

   Price the ordinary and expensive cases

   Use typical calls and the longer calls that dominate the bill. Keep any requests above a model's context threshold in a separate calculation.

2. 2

   Separate interactive work from scheduled work

   Put immediate user requests on the appropriate online tier. Evaluate Batch for report extraction, enrichment and other work with a later deadline.

3. 3

   Add the costs that tokens omit

   Include cache retention, grounding queries, runtime and supporting services. Show promotional effective cost and post-promotion cost separately.

4. 4

   Commit only after measuring demand

   Use observed throughput and the selected model's GSU conversion before purchasing reserved capacity. A cheaper annual unit rate does not fix low utilization.

Vertex AI Cost FAQ

Is Vertex AI now called Agent Platform?

Yes. Google introduced Gemini Enterprise Agent Platform on April 22, 2026 as the evolution of Vertex AI, and its release notes document the naming changes. The developer platform and the Gemini Enterprise employee app remain distinct purchasing concepts.

What are the different price plans for Google Cloud?

Google Cloud has product-specific billing. For the Gemini inference covered here, compare Standard, listed discounted serving options, Priority and Provisioned Throughput. A general Free/Professional/Enterprise app-plan ladder does not describe these model charges.

What is Vertex AI similar to?

For a developer buying model inference, compare Google's Gemini Developer API, Amazon Bedrock and Microsoft Azure. Match the model where possible; otherwise compare cost per accepted result on your own task, with region, tools and supporting services included.

Is Vertex AI pricing the same as Gemini API pricing?

Some Global Standard rates match. Cache storage, cached Batch input, audio rows and free-tier availability can differ. Use the pricing page for the API you will call.

Is Google Vertex free?

Google offers eligible new customers trial credit. The Cloud inference schedule still lists usage prices. Selected Gemini Developer API models have a separate free tier with limits and different data-use terms.

Is there a Vertex AI pricing calculator?

Start with input and billed output token totals multiplied by the model's listed rates, as in the two examples above. Add cached-token processing, token-hour storage and grounding separately. Reserved capacity needs a model-specific throughput estimate as well as a monthly volume estimate.

Is there a student discount?

The cited Cloud generative-AI pricing page does not publish a separate student token rate. Do not apply a consumer Gemini student offer to Cloud API calls without an explicit credit or program entitlement.

What is the refund policy?

The model pricing page does not promise a general refund window for consumed inference. Raise billing discrepancies with Google Cloud support and check the terms attached to your account or promotional credits. Do not build an experiment's budget around an assumed refund.

Will these prices change?

Google currently lists Argon's higher rates from December 1, 2026, and higher 3.8/3.7/3.6 Flash rates from January 1, 2027. Those scheduled changes are already in the budgets above. Recheck the official page before committing spend.

Use the AI Business Workflow Audit Checklist to map each proposed AI job to its volume, deadline and operating requirements before selecting the infrastructure.