Google Vertex AI Pricing 2026: The Complete Cost Breakdown

There is no single price for Google Vertex AI, and that is the first thing to understand about it. Ask what Vertex costs and the honest answer is a dozen meters running at once: tokens for Gemini, node-hours for training, per-hour endpoints that bill whether or not anyone uses them. One team discovered a single forgotten GPU endpoint quietly costing more than $8,000 a month, serving nothing. Vertex is powerful, but its pricing rewards the people who understand the meters and quietly punishes those who do not.

This guide maps every meter. We break down Google Vertex AI pricing in 2026 service by service, Gemini models, training, prediction, search, grounding, and the rest, with current published rates, the fees that tend to hide, and the levers that bring the bill down. The goal is simple: to let you design a Vertex bill on purpose instead of discovering one by accident.

Vertex is usage-based. You pay for what you consume, with three dominant billing units depending on the service: per million tokens for generative models, per node-hour for training and serving infrastructure, and per query or per hour for services like Vertex AI Search and Vector Search. There is no seat fee and no minimum.

The difficulty is rarely a single rate; it is that Google publishes each service's pricing on a separate page with no unified calculator, so estimating a real workload means reading several docs and reconciling them yourself. Two structural facts matter before the numbers. First, some charges keep running even when nothing is in use, which is the root of most Vertex surprises. Second, because Vertex lives inside GCP, committed use discounts, Spot capacity, and regional pricing all apply, so the same job can cost meaningfully different amounts depending on how and where you run it.

Gemini Model Pricing on Vertex AI

For most teams, Gemini model calls are the single largest line on the Vertex bill. Google prices models per million tokens, with separate rates for input and output, and output typically costs several times more than input. The current published rates for the widely used tier look like this.

Newer Gemini 3.x models are available on Vertex at their own, higher rates, and Gemini Pro applies a long-context surcharge, roughly double, once a prompt crosses about 200,000 tokens. Because rates move, confirm current numbers on Google's pricing page before budgeting. For the model lineup and how these rates compare on the direct Gemini API, see our Google Gemini API pricing 2026 guide.

Thinking tokens: the reasoning surcharge

Gemini 3.x models reason internally before answering, and those reasoning, or thinking, tokens are billed as output. A request that returns a short visible answer can quietly consume five to ten times more output tokens than the response you see, which makes reasoning one of the most common sources of unexpected charges. The thinking_level parameter controls it: minimal for simple tasks, medium as a sensible default, and high only when you genuinely need deep reasoning and are prepared to pay for it. Treating that parameter as a cost dial, not a quality default, is one of the easiest Vertex savings.

Context caching, batch, and provisioned throughput

Three mechanics change the effective token price. Context caching serves a repeated prompt prefix, a fixed system prompt or a document you send on every call, from cache at a steep discount, commonly cutting input cost by around 75 percent. Batch prediction processes non-urgent work asynchronously at roughly half price. And provisioned throughput reserves guaranteed model capacity for predictable latency, priced per generative scale unit hour rather than per token, which suits steady, high-volume production traffic. Matching the mode to the workload, cache the repeated context, batch the overnight jobs, provision the always-on chatbot, is where the early savings come from.

Training Costs on Vertex AI

Custom training is the spikiest cost category on Vertex; a single misconfigured distributed job can outspend a month of inference. Training bills per node-hour, combining a machine-hour charge for vCPU and memory with an accelerator charge for any attached GPUs or TPUs, from the moment Google provisions resources until the job completes. A fully loaded custom training node can run upward of $21 an hour, but the accelerator is usually the dominant term.

Prediction and Serving Costs

Once a model is trained or chosen, serving it is where costs quietly accumulate. Vertex offers two prediction modes. Online prediction deploys a model to an endpoint that bills per node-hour continuously while it is deployed, whether or not a single request arrives. Batch prediction runs a job over a dataset and bills only while the job is active, which makes it far cheaper for anything that does not need a live response.

This is the number-one hidden cost on Vertex, and it deserves emphasis: prediction endpoints, Vector Search nodes, and Workbench notebook instances all bill per hour regardless of traffic. That is how a single forgotten H100-backed endpoint reaches roughly $8,100 a month doing nothing. The discipline is blunt but effective: undeploy endpoints the moment a demo or experiment ends, and never leave a GPU-backed node running out of habit.

Vertex AI Search, Agent Builder, and Grounding

Retrieval and agents have their own meters, and they stack on top of token costs. Vertex AI Search, Google's enterprise retrieval and RAG product, costs about $4 per 1,000 search queries plus the underlying tokens. Vertex AI Agent Builder itself carries no incremental fee; you pay for the Gemini tokens the agent consumes, the Vertex AI Search queries it runs, and any tools it calls, such as Cloud Functions or Cloud Run. Grounding a model in live external data adds a per-prompt charge that varies by source.

Other Vertex AI Services to Budget For

Several smaller meters round out a typical bill. Vertex AI Pipelines charge about $0.03 per pipeline run plus the underlying compute. Vector Search bills per node-hour for serving plus an indexing cost. Imagen for image generation and Veo for video generation are priced by output, and video is a notable trap: at roughly $0.50 to $0.75 per second of generated video, it has caused some of the most dramatic billing surprises on the platform. Generative AI Studio, the no-code prompt environment, is free to open, but every prompt you run inside it consumes tokens at standard rates, so prototyping is not free either. And distillation, training a smaller model to mimic a larger one, costs a small per-character training fee and then serves at the cheaper model's price, capturing much of the quality at a fraction of the per-query cost.

The Free Tier and Discounts

New users get a limited Vertex AI free tier, including a monthly allowance of online prediction and some custom training hours at no cost, on top of the standard Google Cloud $300 sign-up credit. Beyond the free tier, the main discount levers are the ones Vertex inherits from Google Cloud: committed use discounts for steady, predictable consumption, Spot capacity for interruptible training and batch work, and regional selection, since the same accelerator can cost noticeably less in one region than another. For how commitment-based discounts work in practice, see our discount manager and cloud pricing models guides.

The Hidden Costs That Inflate a Vertex AI Bill

Most Vertex overspending traces back to a short list of predictable leaks:

• Idle endpoints and nodes. Online prediction endpoints, Vector Search nodes, and Workbench instances bill per hour regardless of traffic. This is the single biggest surprise, and a forgotten GPU endpoint can cost thousands a month for nothing.
• Reasoning tokens. Gemini thinking tokens are billed as output and can multiply a request's cost five to tenfold beyond the visible answer.
• Long-context pricing. Prompts over roughly 200,000 tokens can hit doubled Pro rates, so stuffing context is expensive.
• Video and image generation. Veo's per-second pricing scales fast, and a few long clips add up quickly.
• Egress and region drift. Cross-region traffic and network egress are easy to forget, and running in a pricier region silently raises every rate.

How to Reduce Your Vertex AI Costs

• Tier your models: default to Gemini 2.5 Flash-Lite or Flash for routine work and reserve Pro for genuinely hard reasoning.
• Turn the thinking_level parameter down; use minimal or medium unless a task truly needs deep reasoning.
• Cache repeated context and batch non-urgent jobs; together they can cut token spend dramatically.
• Undeploy idle endpoints immediately, and use Spot VMs for training and batch to save up to 60 to 91 percent.
• Apply committed use discounts to steady workloads, and keep an eye on region and egress.

Because Vertex spend is split across so many meters and sits inside your wider Google Cloud bill, the durable fix is attribution and governance, seeing which team, model, and feature drives each meter, the same discipline we cover in our FinOps for AI and AI cost optimization guides.

The Bottom Line

Vertex AI is not expensive so much as it is plural. The platform hands you a dozen ways to spend and a single invoice to explain them, which is why the teams that stay in control are the ones who treat each meter as a deliberate choice rather than a default. Pick the smallest Gemini model that clears the bar. Run batch where latency does not matter. Undeploy the endpoint the moment the demo ends. Cache the context you send twice, and turn reasoning down when the task is simple.

None of that is exotic. It is just the difference between a Vertex bill you designed and one that happened to you. Read the meters, match each to the job in front of it, and Google Vertex AI's flexibility becomes the advantage it is meant to be rather than a line item nobody can explain at the end of the quarter.