by Google Cloud
Google Vertex AI pricing, plans and limits
Google Cloud's managed AI platform, now called Gemini Enterprise Agent Platform.
- Best for
- ML engineers at Google Cloud shops
- Not for
- Solo builders who want a flat monthly plan
Should you use Google Vertex AI?
Pick it if your company already runs on Google Cloud and wants Gemini, Claude and custom training under one account. Skip it if you need the Pro-tier Gemini under Google's security controls today, because Google says those controls do not cover preview models. Google's uptime objective for Gemini requests is 99.5%.
First-party Gemini, partner models such as Claude and your own GPU or TPU training sit under one Google Cloud account, with agent hosting billed by the vCPU-hour.
First-party Gemini access and a place inside existing Google Cloud accounts give it a strong channel, but Claude and open models are also sold by rival clouds, so the moat is bundling rather than exclusive model access.
Also for
- Platform teams standardising on Gemini and Claude under one bill
- ML engineers who train and serve custom models on GPUs or TPUs
Show 3 more
- Platform teams at companies that already keep their data on Google Cloud
- Agent developers building on Google's Agent Development Kit
- Security leads who need regional data residency and customer-managed keys
Also not for
- Teams that need Google's security controls on preview-stage models
Main trade-off
Names and documentation are mid-migration after the April 2026 rename, and the newest Gemini Pro tier is still in preview.
Google Vertex AI pricing
Server plans
| Plan | GPU | Price |
|---|---|---|
| NVIDIA T4 acceleratorGPU · Custom training accelerator in us-central1, billed on top of the machine type. Small models and prototypes. | 1× NVIDIA T4 | $0.40/hr |
| NVIDIA L4 acceleratorGPU · Custom training accelerator in us-central1, billed on top of the machine type. | 1× NVIDIA L4 | $0.64/hr |
| NVIDIA A100 acceleratorGPU · Custom training accelerator in us-central1; a management fee of $0.44 an hour applies on top. | 1× NVIDIA A100 | $2.93/hr |
| NVIDIA A100 80GB acceleratorGPU · Custom training accelerator in us-central1; a management fee of $0.59 an hour applies on top. | 1× NVIDIA A100 80GB, 80 GB VRAM | $3.93/hr |
| NVIDIA H100 80GB acceleratorGPU · Large model training accelerator in us-central1; a management fee of $1.47 an hour applies on top. | 1× NVIDIA H100 80GB, 80 GB VRAM | $9.80/hr |
| NVIDIA H200 141GB acceleratorGPU · Large model training accelerator in us-central1, billed on top of the machine type. | 1× NVIDIA H200 141GB, 141 GB VRAM | $10.71/hr |
| TPU 7x (tpu7x-standard-4t)GPU · Google's own accelerator for custom training in us-central1. | 1× Google TPU 7x, 1 chip | $13.80/hr |
| Gemini 4 Argon (Global)Model API · $2 input / $10 output per 1M tokens through 30 November 2026, then $4 / $20 from 1 December 2026; cached input $0.10, then $0.20. · Global endpoint | — | — |
| Gemini 3.1 Pro Preview (Global)Model API · $2 input / $12 output per 1M tokens up to 200K input tokens; $4 / $18 above that. Priced as a Preview model. · Global endpoint | — | — |
| Gemini 3.8 Flash (Global)Model API · $0.75 input / $3.75 output per 1M tokens through 31 December 2026, then $1.50 / $7.50 from 1 January 2027. · Global endpoint | — | — |
| Gemini 3.5 Flash (Global)Model API · $1.50 input / $9 output per 1M tokens; cached input $0.15. · Global endpoint | — | — |
| Gemini 3.5 Flash-Lite (Global)Model API · $0.30 input / $2.50 output per 1M tokens, for high-volume routing and extraction. · Global endpoint | — | — |
| Gemini 3.1 Flash-Lite (Global)Model API · $0.25 input / $1.50 output per 1M tokens for text, image and video input; audio input $0.50. · Global endpoint | — | — |
| Gemma 4 26BModel API · $0.15 input / $0.60 output per 1M tokens, cache hit $0.015. Google's open-weights model, served as a managed API. · Model Garden | — | — |
| Claude Sonnet 5.5 (Global)Model API · $2 input / $10 output per 1M tokens; batch $1 / $5; cache hit $0.10. · Global endpoint | — | — |
| Claude Opus 5.5 (Global)Model API · $4 input / $20 output per 1M tokens; batch $2.50 / $12.50; cache hit $0.20. · Global endpoint | — | — |
| Claude Haiku 5.5Model API · $0.10 input / $0.50 output per 1M tokens for prompts up to 100K tokens; $0.50 / $2.50 above that; batch $0.05 / $0.25. · Global endpoint | — | — |
| Provisioned Throughput (Global, per GSU)Cloud · Reserved Gemini capacity billed per generative AI scale unit and hour: $7.14 on a 1 week commit, $3.70 on 1 month, $3.29 on 3 months, $2.74 on 1 year. · Global endpoint | — | — |
| Agent Runtime (per vCPU-hour)Cloud · $0.085 per vCPU-hour after 50 free hours a month, plus $0.009 per GiB-hour of memory after 100 free GiB-hours. Idle time between turns is not billed. · Multi-region and regional | — | — |
- Minimum commitment
- None on demand; Provisioned Throughput commits run 1 week, 1 month, 3 months or 1 year
- Billing
- Pay as you go · Committed spend
How the pricing works, in full
Per 1M tokens (input / output, standard prompt size, global endpoint) from Google's pricing page checked on 2026-10-26: Gemini 4 Argon $2 / $10 through 30 November 2026 and $4 / $20 from 1 December 2026, Gemini 3.1 Pro Preview $2 / $12, Gemini 3.5 Flash $1.50 / $9, Gemini 3.5 Flash-Lite $0.30 / $2.50 and Gemini 3.1 Flash-Lite $0.25 / $1.50. Gemini 3.8 Flash is $0.75 / $3.75 through 31 December 2026 and $1.50 / $7.50 from 1 January 2027, so 80M input and 8M output tokens a month cost $90 now and $180 after the change. Regional endpoints add 10% for generally available Gemini 3 and later models. Claude on this platform (Claude Haiku 5.5 $0.10 / $0.50 under 100K input tokens, Claude Sonnet 5.5 $2 / $10, Claude Opus 5.5 $4 / $20) carries the same global rates as Claude in Amazon Bedrock. Custom training bills accelerators by the hour in us-central1, from $0.4025 for a T4 to $9.80 for an H100 80GB, in 30-second increments. Provisioned Throughput is sold per generative AI scale unit and hour, from $7.14 on a 1 week commit to $2.74 on a 1 year commit. Agent Runtime bills $0.085 per vCPU-hour after the first 50 free hours each month.
Costs that are easy to miss
- Grounding with Google Search is free for the first 5,000 queries a month across Gemini 3 models, then $14 per 1,000 queries.
- Tuned Gemini 3 endpoints predict at 1.5 times the base model price.
- A100 and H100 custom training adds a per-accelerator management fee on top of the accelerator rate.
Google Vertex AI regions and compliance
- Uptime SLA
- 99.5% for Gemini requests; 99.9% for training, deployment and batch prediction
Available regions
- Columbus, Ohio (us-east5)
- Dallas, Texas (us-south1)
- Iowa (us-central1)
- Las Vegas, Nevada (us-west4)
- Los Angeles, California (us-west2)
- Moncks Corner, South Carolina (us-east1)
- Northern Virginia (us-east4)
- Oregon (us-west1)
- Salt Lake City, Utah (us-west3)
- Montreal (northamerica-northeast1)
- Toronto (northamerica-northeast2)
- Santiago, Chile (southamerica-west1)
- Sao Paulo, Brazil (southamerica-east1)
- Johannesburg, South Africa (africa-south1)
- Belgium (europe-west1)
- Finland (europe-north1)
- Frankfurt, Germany (europe-west3)
- London, United Kingdom (europe-west2)
- Madrid, Spain (europe-southwest1)
- Milan, Italy (europe-west8)
- Netherlands (europe-west4)
- Paris, France (europe-west9)
- Turin, Italy (europe-west12)
- Warsaw, Poland (europe-central2)
- Zurich, Switzerland (europe-west6)
- Hong Kong (asia-east2)
- Jakarta, Indonesia (asia-southeast2)
- Melbourne, Australia (australia-southeast2)
- Mumbai, India (asia-south1)
- Delhi, India (asia-south2)
- Osaka, Japan (asia-northeast2)
- Seoul, Korea (asia-northeast3)
- Singapore (asia-southeast1)
- Sydney, Australia (australia-southeast1)
- Taiwan (asia-east1)
- Tokyo, Japan (asia-northeast1)
- Dammam, Saudi Arabia (me-central2)
- Doha, Qatar (me-central1)
- Tel Aviv (me-west1)
Compliance certifications
- ISO 27001
- SOC 2
- HIPAA eligible
As stated by the provider. Ask for the current report before you rely on one.
Managed services
- Model Garden
- Agent Studio
- Agent Runtime
- Memory Bank
- Agent Gateway
- Provisioned Throughput
- Vector Search
- Model Registry
- Feature Store
- Pipelines
- AutoML
Access methods
- Web console
- API
- CLI
- Dashboard
What Google Vertex AI does
Google Vertex AI, sold since April 2026 as Gemini Enterprise Agent Platform, is Google Cloud's managed AI platform, and Google says its Model Garden holds more than 200 models. It serves Gemini, Claude and open models through one API, trains custom models on GPUs and TPUs, and hosts long-running agents.
Model Garden with first-party and partner models
Model Garden serves Gemini and Gemma, and Google's pricing page lists Gemini and Gemma alongside partner models from Anthropic, xAI, DeepSeek, MiniMax, Moonshot, Qwen, GLM, OpenAI, Meta and Mistral.
Regional and global endpoints
Gemini 3 and later run on a global endpoint or on regional endpoints, machine learning services sit in 39 named regions with no global location, and Google says it stores and processes data only in the region you choose, apart from data labeling and preview features.
Provisioned Throughput
Reserved Gemini capacity is sold in generative AI scale units on terms from 1 week to 1 year, backed by a 99% latency target for streaming requests under Google's Provisioned Throughput SLA.
Agent Runtime and Memory Bank
Hosts agents that run for up to 7 days, according to Google's April 2026 release notes, with sub-second cold starts, persistent memory and sessions, and no charge for idle time between turns.
Show all 6 features
GPU and TPU training with Spot VMs
Custom training runs on T4, L4, A100, H100 and H200 GPUs or TPUs, billed by the hour in 30-second increments, and it accepts Spot VMs and Compute Engine reservations.
Four listed security controls
Google's generative AI security page lists data residency at rest, customer-managed encryption keys, VPC Service Controls and Access Transparency, and its HIPAA page names the platform among Business Associate Agreement products.
Strengths and limits
Strengths
- One Google Cloud account covers Gemini, Claude and open models, so swapping the model behind an application does not need a second vendor contract.
- Hourly training rates are published with 1 year and 3 year savings-plan columns, which lets a team cost a training run before launching it.
- Requests that fail with a 4xx or 5xx code are not billed for tokens, so a retry storm during an outage does not inflate the invoice.
- Google publishes uptime objectives with credit schedules for both platform services and Gemini inference, which gives procurement something to hold the vendor to.
Limits
- Gemini 3.1 Pro is priced as a Preview model, and Google states that its security controls are not supported for preview models, so regulated workloads cannot use it under those controls.
- The April 2026 rename left names inconsistent: pricing and compliance pages say Agent Platform, while the API host is still aiplatform.googleapis.com and many documentation and console paths still read vertex-ai, and guides written earlier use retired product names.
- An endpoint that hosts your own model bills node-hours while it waits for traffic, and Google says scale-to-zero is not supported, so an idle endpoint still costs money until you undeploy it.
- Two launch discounts expire soon: Gemini 4 Argon's rate doubles on 1 December 2026 and Gemini 3.8 Flash's on 1 January 2027, so a budget built on today's token prices will run short.
Google Vertex AI alternatives
| Service | Regions | Pricing | Compliance | Checked |
|---|---|---|---|---|
| Google Vertex AI This page | US · Canada · South America · Africa +3 | Free tier · Pay as you go · Committed spend | ISO 27001 · SOC 2 · HIPAA | 26 Oct |
 Claude in Amazon BedrockPick Vertex if Gemini or custom GPU training is in the plan; pick Bedrock if AWS must be the only data processor. | Africa · Asia Pacific · Canada · Europe +3 | Pay as you go · Committed spend | HIPAA · GDPR · FedRAMP | 10 Oct |
 AnthropicPick Anthropic's own API if Claude is the only model you need; pick Vertex if you also want Gemini and one Google Cloud bill. | US · Europe · Asia Pacific | Free tier · Pay as you go · Subscription | SOC 2 · ISO 27001 · ISO 42001 +1 | 9 Oct |
 LuminalCompiler-built inference for open models, as a serverless API or on-prem. | -- | Pay as you go | -- | 23 Oct |
Google Vertex AI facts
- Provider founded
- 2021
- Headquarters
- Mountain View, CA, USA
- Open source
- No
What changed
Questions about Google Vertex AI
How much do you pay for Google Vertex AI?
You pay per use, with no subscription: Google's pricing page (checked 26 October 2026) lists Gemini 3.5 Flash-Lite at $0.30 input and $2.50 output per 1M tokens and Gemini 4 Argon at $2 and $10, with Anthropic's Claude models at the same global rates Amazon Bedrock lists. Training and serving bill node-hours, agents bill vCPU-hours. A support chatbot that sends 2,000 input and 300 output tokens per reply to Gemini 3.5 Flash-Lite costs about $13.50 per 10,000 replies. Google's 1 year and 3 year Flexible Savings Plans cut the published hourly rates for training and agents by 10% and 20%.
What do you get on Google Vertex AI's free tier?
New Google Cloud customers get up to $300 in free credits, valid for 90 days, according to Google's product page. Beyond the trial, the free allowances are small: 5,000 grounded search queries a month across Gemini 3 and Gemini 4 models, plus the first 50 vCPU-hours of agent compute each month per account. The pricing page lists no free monthly token allowance for Gemini models.
What are Google Vertex AI's closest competitors?
Claude in Amazon Bedrock is the nearest match when your data already sits in AWS, and it lists the current Sonnet and Opus models at the same global rates. Anthropic's own API suits teams that need only Claude and want no cloud contract in between.
How does Google Vertex AI compare to Claude in Amazon Bedrock in 2026?
For the current Sonnet and Opus models the two list identical global token rates, so cost does not decide it. Vertex adds first-party Gemini and your own GPU and TPU training, while Anthropic's documentation says Bedrock's Claude endpoint runs on AWS-operated infrastructure that Anthropic staff cannot access. Pick Vertex when Gemini or custom training is in the plan, and Bedrock when AWS must be the only data processor.
What does it take to start using Google Vertex AI?
Sign in to Google Cloud, start the free trial if you are new, and open Agent Platform at console.cloud.google.com/agent-platform/overview. Pick a region first, because machine learning resources are tied to one location and API calls go to a regional host such as us-central1-aiplatform.googleapis.com. Google requires training data and program files to sit in Cloud Storage, and it warns that a large cluster on a first job is likely to exceed your quota, so start small.
About Google Vertex AI
Google Vertex AI is the managed AI platform of Google's cloud division. Google Cloud announced its general availability on 18 May 2021, and on 22 April 2026 its release notes renamed it: Vertex AI Platform became Gemini Enterprise Agent Platform, Vertex AI Studio became Agent Studio and Agent Engine became Agent Runtime. Google's announcement the next day calls the new platform the evolution of Vertex AI and says all Vertex AI services and roadmap work will be delivered only through it. This record keeps the Vertex AI name so that people who know the product by it can find it, and it follows Google's current documentation for every fact below.
Read the full background
The platform sells three separate jobs. The first is model access: Gemini models run on Google's own endpoints, and Model Garden adds partner models such as Claude plus open models such as Gemma 4. The second is custom machine learning: you train and serve your own models on GPUs, TPUs or plain CPUs, with AutoML for image and tabular data. The third is agents: Agent Runtime, Memory Bank, sessions and a code sandbox host long-running agents built with Google's Agent Development Kit.
The natural buyer is a team that already keeps its data on Google Cloud. Gemini 4 Argon, Google's new frontier model, became generally available on 13 October 2026, Claude Haiku 5.5 reached Model Garden on 7 October, and 16 older open-model endpoints were retired on 21 October, so the model list moves fast. Claude is also sold by Anthropic directly and through Claude in Amazon Bedrock, which makes the deciding question where your other systems live, not which cloud has the model. Browse the wider field under inference hosting and GPU compute, or let Smart Match narrow the choice for your workload.
Sources and guides
Official sources
- Google Vertex AI official site Vendor
- GitHub repository GitHub
Plans, regions, the uptime commitment and the compliance list on this page are read from Google Cloud's own pages and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Hokai guides that mention Google Vertex AI
Lists that include Google Vertex AI
More AI Infrastructure Services on Hokai
PineconeManaged vector database for search, RAG and agent memory.
AgentMailEmail inbox API built so every AI agent can own, read and answer its own inbox- Amazon SESAWS email sending and receiving, metered per thousand messages
Cloudflare Email ServiceSend and receive email from Workers and agents, with unlimited inbound routing
PostmarkTransactional email API that keeps app mail apart from bulk sendsResendEmail API for developers, with inboxes your AI agents can read and answer