Vertex AI pricing for the Gemini models is per million tokens, input and output at different rates per model, with cached input at a tenth of the input rate, a higher rate above 200,000 tokens of context on the Pro models, audio input at its own rate, and batch at half price; provisioned throughput replaces token billing with generative AI scale units per hour over a week, a month, three months or a year; context cache storage costs per million token-hours; tuning per training token and hosting a tuned model per hour; custom models pay node-hours for training and prediction endpoints, Vector Search node-hours and index storage, and Cloud Storage and BigQuery for what the models read and write. On the bill, output tokens dominate on chat and generation, because they cost four to six times input, and prediction endpoints dominate on custom models, because a node pays its hour whether requests come or not.
Pricing dimensions
Google prices a foundation model by the tokens that pass through it and a custom model by the machine that serves it; every model, modality, tier and endpoint has its own SKU, and the global endpoint carries the base rate, with regional endpoints about ten percent above it for the Gemini 3 models from 2026-07-01.
| Dimension | Unit | What drives it | Optional |
|---|---|---|---|
| Gemini input tokens | Million tokens of text, image, video or audio per model; the Pro models at a higher rate above 200,000 tokens of context; audio at its own rate on the Flash models | Prompt, system instruction, history, documents and media on every call | No |
| Gemini output tokens | Million tokens of response and reasoning per model, several times the input rate | Every answer, thinking tokens included | No |
| Cached input | Million cached tokens at a tenth of the input rate, plus context cache storage per million token-hours for explicit caches | Repeated prefixes: system prompts, documents, few-shot examples | Yes |
| Service tiers | Standard; Priority at a higher rate for precedence; Flex or Batch at half the rate for asynchronous jobs | The latency the application can accept | Yes |
| Provisioned throughput | Generative AI scale units per hour, committed for a week, a month, three months or a year at falling rates; global or regional | Reserved capacity for a model, used or not | Yes |
| Tuning and tuned models | Training tokens per million; the tuned model's inference at the base model's token rates on Gemini | Supervised tuning jobs | Yes |
| Custom training and prediction | Node-hours by machine type with a Vertex AI premium over Compute Engine, GPUs and TPUs per hour; prediction endpoints per node-hour around the clock | Custom models, AutoML, endpoints with minimum replicas | Yes |
| Vector Search | Node-hours of the serving index plus GiB-month of index storage | Embedding search at scale | Yes |
| Embeddings and other models | Per 1,000 characters or per million tokens for embeddings; per image, per second of video or audio for the media models; partner models at their own rates | Retrieval pipelines, media generation, third-party models | Yes |
| Around the models | Cloud Storage for datasets and artifacts, BigQuery for data, Cloud Logging for request logging | Every pipeline's storage | Yes (Cloud Storage pricing, BigQuery pricing) |
| Purchase option | Provisioned throughput commitments; the Flexible Savings Plan across eligible products; nothing on tokens otherwise | Steady inference load | Yes (Google Cloud committed use discounts) |
The model is the price: a Flash model costs a fraction of a Pro model per token, and within a model the output token costs four to six times the input, so a chat that produces long answers pays mostly for them, and an extraction task that reads much and writes little pays mostly for input. Context is input on every call, the whole history, system prompt and documents anew, which is why the Pro models raise the rate above 200,000 tokens and why caching pays: a cached prefix costs a tenth per token plus a storage fee per hour it lives, so a document asked ten questions in an hour costs less cached than resent. Batch halves both rates for anything that can wait, and provisioned throughput turns the token price into a reservation per scale unit and hour, which pays off only when the units are busy, because they cost whether tokens flow or not. Custom models are the opposite economy: an endpoint pays its node every hour and the first request as much as the millionth, so a model served to few requests costs more than the same work on a Gemini model per token. The mechanics across all three providers are in AI and machine learning services; what the model's rate card means, in model pricing and rate cards; what the AWS counterpart costs, in Bedrock and SageMaker pricing.
How it appears in your billing data
In the billing export the rows carry service.description Vertex AI and a sku.description per model, modality and kind: Gemini 2.5 Pro Text Input Tokens, Gemini 2.5 Pro Text Output Tokens, Gemini 2.5 Flash Cached Input Tokens, Gemini 2.5 Flash Audio Input Tokens, Gemini 2.5 Pro Batch Text Input Tokens for batch, Gemini 2.5 Pro Long Context Text Input Tokens for the rate above 200,000, Generative AI Scale Unit per hour for provisioned throughput, Context Cache Storage per token-hour, Vertex AI Online Prediction with the machine type and region for endpoints, Vertex AI Training for jobs, Vector Search Serving Node; usage.unit is count for tokens, hour for units and nodes. resource.name in the export with resource-level detail names the endpoint or the tuning job, but for the Gemini models nothing that tells applications apart: all tokens of a model from one project sum into one row, and attribution to applications, customers or features needs a project per application, the request's labels where the API accepts them, or the application's own token count per customer, as AI cost allocation describes.
Discounts sit in credits: a provisioned throughput commitment or a Flexible Savings Plan appears as COMMITTED_USAGE_DISCOUNT or FEE_UTILIZATION_OFFSET beside the rows it covers, and promotional credits for new models as PROMOTION. The export's structure is described in the Google Cloud billing export; in the FOCUS export, ServiceName carries Vertex AI, ServiceCategory AI and Machine Learning, SkuId the SKU, PricingCategory Standard or Committed and CommitmentDiscountId the commitment (FOCUS columns). What the export does not carry is the request count or the tokens per request; the Vertex AI metrics and the request logs, when enabled, hold them.
Levers
Configuration: the first lever is the model per task, because a Flash model costs a fraction of a Pro model and suffices for classification, extraction and summarisation; the second is the prompt, because every call pays the whole context and a trimmed history, a shorter system prompt and fewer examples cut the input on every call; the third is caching for repeated prefixes and batch for everything that can wait, as caching, batching and routing describes. Output length is the fourth: a maximum output token limit and a format that asks for the answer, not the essay, halve the dearest tokens (tokens, requests and context).
Purchase option: provisioned throughput for a steady, latency-sensitive load whose scale units stay busy, over the term the load is certain for; the Flexible Savings Plan where the spend spans services; on-demand tokens for everything bursty, because an idle scale unit costs its hour (commitment strategy).
Architecture: Gemini models for anything a foundation model does, custom endpoints only for models that cannot be a prompt, and those with autoscaling to zero replicas where the service allows it or batch prediction where it does not; embeddings computed once and stored, not recomputed per query; and the global endpoint where data residency allows it, because the regional endpoint costs more and offers less capacity. What the GPUs behind custom training cost on Compute Engine is in Compute Engine pricing.
Where it goes wrong
- The whole history on every turn. A chat that resends the conversation pays it as input every call and crosses 200,000 tokens on the Pro model, where the rate rises. Summarise the history and cache the system prompt.
- Pro for the simple task. Classification on a Pro model pays several times the Flash price for the same label. Route by task; the Flash model is the default.
- The thinking tokens are output. Reasoning tokens are billed as output at the output rate. A thinking budget on the request caps them.
- Provisioned throughput at 20 percent. Scale units pay their hour whether tokens flow or not; below about half utilisation on-demand is cheaper. The token line against the units' capacity shows it.
- The endpoint serves ten requests a day. A prediction endpoint with one node pays 730 node-hours a month for minutes of work. Batch prediction or a Gemini call replaces it.
- The cache that nobody hits. An explicit context cache pays storage per token-hour; a cache created per request and never reused pays storage for nothing. Cache what repeats, with a TTL that matches the session.
- One project for every application. All tokens of a model fall into one row. A project per application, or the application's own token count, is the only attribution.
Example
List prices in USD on the global endpoint, read on 2026-10-21 from the Vertex AI generative AI pricing page, the same for a request from Frankfurt (europe-west3) as from anywhere: Gemini 2.5 Flash costs USD 0.30 per million input tokens, USD 0.03 per million cached input tokens, USD 1.00 per million audio input tokens and USD 2.50 per million output tokens; Gemini 2.5 Flash Lite USD 0.10, USD 0.01 and USD 0.40; Gemini 2.5 Pro USD 1.25 per million input tokens up to 200,000 tokens of context and USD 2.50 above, USD 0.125 and USD 0.25 cached, USD 10.00 and USD 15.00 per million output tokens; Gemini 3.1 Pro Preview USD 2.00 and USD 4.00 input, USD 0.20 and USD 0.40 cached, USD 12.00 and USD 18.00 output; Gemini 3.5 Flash USD 1.50 input, USD 0.15 cached and USD 9.00 output; batch at half the rates; context cache storage USD 4.50 per million token-hours on the Pro models and USD 1.00 on the Flash models; provisioned throughput USD 7.14 per scale unit-hour for a one-week commitment, USD 3.70 for a month, USD 3.29 for three months and USD 2.74 for a year.
An application that sends 10 million input tokens and 2 million output tokens a month pays on Gemini 2.5 Flash 3.00 plus 5.00 = USD 8.00, on Gemini 2.5 Pro 12.50 plus 20.00 = USD 32.50, on Gemini 3.1 Pro 20.00 plus 24.00 = USD 44.00; in batch half of each. The same load on 2.5 Pro with 60 percent of the input served from cache pays 5.00 plus 0.75 plus 20.00 = USD 25.75 plus the cache's storage, a 20,000-token system prompt kept for 730 hours USD 0.07. A chat that resends a 150,000-token history on every one of 10,000 turns a month pays 1.5 billion input tokens, USD 1,875 on 2.5 Pro, and USD 187.50 with the history cached. One scale unit of provisioned throughput costs USD 2,700 a month on a one-month commitment and USD 2,000 on a one-year commitment, whether tokens flow or not. The figures show that output tokens carry most of a chat's bill, that the model choice moves the bill fivefold and caching threefold, and that a reservation needs a busy month to beat the tokens.
Sources
Read on 2026-10-21. Prices: Vertex AI generative AI pricing, Vertex AI pricing. Service: context caching overview, provisioned throughput, batch prediction for Gemini. Billing data: structure of the standard Cloud Billing data export.