Skip to main content

Decision brief

Gemini Enterprise Agent Platform Review: Buying Gemini and Partner Models Through Google Cloud

Agent Platform, formerly Vertex AI, sells Gemini, Claude and open models inside a Google Cloud project. Check spend-based throughput tiers, endpoint residency and data settings before production.

Reading time
4 min read
Sources
18 cited
In this piece
Gemini Enterprise Agent Platform−27 Negative

GILD verdict

Standardize on Gemini Enterprise Agent Platform when your organization runs on Google Cloud and wants Gemini and Claude under its existing agreement, IAM and residency controls.

Global endpoint prices match each vendor's list price as of Oct 28, 2026. Slack and Fin run it as one of two or three model clouds, so keep a second route ready. Before production traffic, plan throughput for the spend tier you will actually reach, test capacity in your region, restrict every API key, and check every Claude API feature you depend on.

Gemini Enterprise Agent Platform

AI Model API by Google

−27 Negative
Product page

Fit and limits

Who it fits

  • Your organization runs on Google Cloud, and a new model vendor contract or security review would take months.
  • You want Gemini and Claude, plus open models such as Gemma, Llama, DeepSeek and gpt-oss-120b, behind one IAM policy and one bill.
  • Processing must stay in the US or the EU: the us and eu multi-region endpoints keep requests inside that boundary.
  • You want managed agent building blocks in the same project: the Agent Development Kit, Agent Runtime with Memory Bank, and Agent Identity, Registry and Gateway.

Where it falls short

  • OpenAI's GPT-6 models are not in Model Garden; gpt-oss-120b and gpt-oss-20b are the OpenAI models on the pricing page.
  • Google-managed Claude lacks Anthropic's Message Batches API, Files API, server-side code execution and web fetch, Agent Skills and the MCP connector.
  • Standard PayGo throughput grows with 30-day spend, so a new organization starts at Tier 1, and public accounts describe persistent 429 errors after moving from AI Studio.
  • Spend has no owner without labels: one composite postmortem traced a tripled overnight bill to a retry loop only after a two-hour investigation.
  • Keys are the largest bill risk. A key Firebase Hosting published ran 131,950 Vertex AI image calls, US$6,562 against a US$25 budget, at MOVILIXA (Aug 2026), and a stolen service-account key ran up about US$195k on Claude overnight while billing alerts lagged (Jun 2026).
  • Capacity outside the US can run short: a UK team saw up to 80% of Gemini requests return 429 errors on the London, EU and global endpoints for over a week, and none on the us endpoint (Oct 2026).

Implementation reality

  1. Choose models and endpoints

    Run your own test set on two or three Model Garden models through the endpoint you will use in production. The global endpoint carries the list price for Gemini 3 and later models, and every non-global endpoint costs 10% more. Choose the us or eu multi-region endpoint when residency rules require it.

  2. Plan throughput before launch

    Standard PayGo has no fixed per-model quota. An organization's 30-day spend sets its tier, from 500,000 tokens per minute at Tier 1, for US$10 to US$250 of spend, to 10,000,000 at Tier 4, for over US$50,000, on Gemini Pro and Argon models. Traffic above the tier can return 429 errors. For steady traffic, Provisioned Throughput reserves GSUs on 1-week to 1-year terms.

  3. Set data settings

    Google does not train on customer data without permission on managed models. Zero data retention needs two actions: opt out of abuse-monitoring prompt logging and turn off caching. Check the security controls matrix for each feature, because CMEK, VPC Service Controls and data residency at rest are listed per feature.

  4. Label and cap spend

    Attach labels to requests so Cloud Billing can split spend by service and feature, and set budgets with alerts on each project before launch. Grounding with Google Search bills per request on top of tokens, which one solo developer found only after his bill passed about four times his token telemetry.

What to verify before you buy

Models and throughput

  • Standard PayGo throughput: no fixed per-model quota. 30-day organization spend sets the tier, from 500,000 tokens per minute at Tier 1 to 10,000,000 at Tier 4 for Gemini Pro and Argon.
  • Your own model: pre-GA custom weights import deploys fine-tuned supported base models such as Llama and Gemma from a Cloud Storage bucket in the project.

Identity, audit and compliance

  • Single sign-on: Workforce Identity Federation gives users in your external identity provider single sign-on to Google Cloud resources.
  • Audit logs: Admin Activity logs are always on. Data Access logs are off by default and not written unless you enable them.
  • Compliance: Google's table lists Generative AI on Gemini Enterprise Agent Platform (formerly Generative AI on Vertex AI) in scope for ISO 27001, 27017 and 27018 certifications, SOC 1, 2 and 3 reports, PCI DSS certification and penetration testing.

Data use and retention

  • Training: Google says it won't use your data to train or fine-tune any AI or ML models without your prior permission or instruction, on all managed Agent Platform models, GA and pre-GA.
  • Retention: Google says Models as a Service customer data is retained for limited periods in six areas: abuse-monitoring prompt logging for Google models, abuse-monitoring prompt and response logging for Advanced AI models, grounding with Google Search, grounding with Google Maps, request-response logging and the Interactions API.
  • Zero data retention: requires specific customer action in each of those six areas. Google says it may not be possible with some Advanced AI features, and grounding storage can't be disabled if you use grounding with Google Search or Google Maps.

What to measure in your pilot

  • Recheck the pricing page on Nov 30 and Dec 31, 2026, when the Gemini 4 Argon and Gemini 3.x Flash promotions end, and rebudget Argon at US$4 and US$20 per million tokens.
  • GILD found no published turnaround for a custom throughput tier; ask your Google account team before launch and record how long it takes.
  • Measure your own 429 rate on the global endpoint for two weeks before choosing between Standard PayGo and Provisioned Throughput.

Sources

Official documentation
18
  1. Agent Platform documentation: OverviewOfficial documentationdocs.cloud.google.com
  2. Google Cloud: Gemini Enterprise Agent PlatformOfficial documentationcloud.google.com
  3. Google Cloud blog: Introducing Gemini Enterprise Agent Platform (Apr 22, 2026)Official documentationcloud.google.com
  4. Agent Platform documentation: Name changes from Vertex AIOfficial documentationdocs.cloud.google.com
  5. Google Cloud: Agent Platform generative AI pricingOfficial documentationcloud.google.com
  6. Agent Platform documentation: Standard PayGoOfficial documentationdocs.cloud.google.com
  7. Agent Platform documentation: Provisioned ThroughputOfficial documentationdocs.cloud.google.com
  8. Agent Platform documentation: Deployment and endpoint locationsOfficial documentationdocs.cloud.google.com
  9. Agent Platform documentation: Agent Platform and zero data retentionOfficial documentationdocs.cloud.google.com
  10. Agent Platform documentation: Security controlsOfficial documentationdocs.cloud.google.com
  11. Google Cloud: Agent Platform audit loggingOfficial documentationcloud.google.com
  12. Google Cloud: Services in scope for compliance programsOfficial documentationcloud.google.com
  13. Google Cloud: Gemini online inference SLAOfficial documentationcloud.google.com
  14. Agent Platform documentation: Deploy models with custom weightsOfficial documentationdocs.cloud.google.com
  15. Google Cloud IAM: Identity federation: products and limitationsOfficial documentationdocs.cloud.google.com
  16. Google Cloud IAM: Workforce Identity FederationOfficial documentationdocs.cloud.google.com
  17. Google Cloud: Agent Platform audit logging informationOfficial documentationdocs.cloud.google.com
  18. Agent Platform docs: Overview of Model GardenOfficial documentationdocs.cloud.google.com

Keep reading

  1. Buyer guideChoosing Between an AI Studio Key and Google Cloud's Agent PlatformGemini Developer API and Enterprise Agent Platform compared on setup, billing, limits, regions, and data terms.
  2. Buyer guideAgent Platform Production PlaybookA checklist for taking a Gemini or Claude workload on Google Cloud from prototype to live traffic: models and endpoints, throughput, data settings and spend controls.
  3. Buyer guideAgent Platform Data, Endpoints & Residency ChecklistWhich endpoint keeps processing in the US or EU, what zero data retention requires, and which controls cover which models.
Read the research