Gemini Enterprise Agent Platform
The first passport to score all twenty elements: Google renamed Vertex AI on 22 April and now sells the widest agent surface on this table, from embeddings to A2A. The bill is the problem — one question can touch four meters, and nobody quotes you a platform price.
PRICING
| Free trial | $300 | new Google Cloud accounts, 90 days, any GCP service |
| Express Mode | $0 | Agent Studio and Agent Platform with limited quotas, up to 10 agent engines, 90 days, no billing account |
| Agent Runtime | $0.0864 / vCPU-hr | plus $0.0090 per GB-hour; 50 vCPU-hours and 100 GB-hours free monthly; Code Execution billed at the same rates |
| Sessions & Memory Bank | $0.25 / 1,000 | stored events or memories; retrieved memories $0.50 per 1,000; LLM costs billed separately |
| Agent Search | $4.00 / 1,000 | queries; $6.00 with advanced LLM; 10,000 free per month |
There is no subscription and no seat price to compare. A single user question can bill four ways: search or grounding on retrieval, model tokens on generation, a session event on state, and runtime vCPU-hours on execution. Rates were verified on the live pricing page on 05/09/2026. Two dates are worth flagging before budgeting: the live page says Sessions and Memory Bank billing 'will commence on September 1st, 2026', while a Google blog from December 2025 announced the same SKUs for 28 January 2026 — the date moved, so check it again rather than trusting either post. Semantic Governance Policy billing is still 'later in 2026' and prices at one vCPU-hour per 15,000 response evaluations plus evaluation-model tokens. Runtime rates fell in December 2025 ($0.0994 to $0.0864 per vCPU-hour) but Code Execution, Sessions and Memory Bank all moved from free to metered in the same round, so the direction of travel is more meters, not fewer.
checked 2026-09-05 · vendor pricing page
Element scores
Strengths
Twenty of twenty elements applicable, the first passport on this table to cover the whole periodic table, and nothing on it is a stub. Model choice runs from Gemini 3.1 Pro to Claude Opus to open Gemma across 200+ entries in Model Garden. Retrieval has three real paths (RAG Engine, Agent Search, Google Search grounding) sitting next to a production vector service. ADK is open source, stable in five languages, and deploys to Agent Runtime, Cloud Run or your own containers with one command, so the framework is not a lock-in point even when the runtime is. Protocol support is the best on the table: Google wrote A2A, gave it to the Linux Foundation, and built a gateway that inspects both A2A and MCP traffic as a first-class network function. Gen AI evals with trajectory metrics is genuine regression testing for agents, which most of this table still cannot do at all.
Honest dings
The bill is the risk, not the capability. Four to six meters per interaction, model tokens on top, no platform price to anchor a budget, and independent reviewers report surprise invoices running into five figures from services teams forgot were running. The 22 April rename touched almost every product name at once — Vertex AI Studio became Agent Studio, Agent Engine became Agent Runtime, Vertex AI Search became Agent Search — so every tutorial, blog post and Stack Overflow answer written before that date now uses names the console no longer shows, while Google's own docs still serve pages titled 'Vertex AI Agent Builder'. Governance is oversold relative to what has shipped: agent identity is preview and Agentic IAM is roadmap. And all of it assumes Google Cloud; outside GCP the identity, networking and data layers that make this coherent are somebody else's problem.
Sources (16) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- cloud.google.com/blog/products/ai-machine-learning/… — Official launch post 22/04/2026: Agent Platform is the evolution of Vertex AI, and 'all Vertex AI services and roadmap evolutions will be delivered exclusively through the Agent Platform, rather than as a standalone service' (accessed 2026-09-05)
- docs.cloud.google.com/gemini-enterprise-agent-platf… — Official release notes 22/04/2026 carrying the full Vertex-to-Agent-Platform naming table: Vertex AI Studio to Agent Studio, Agent Engine to Agent Runtime, Vertex AI Search to Agent Search, Gen AI evaluation service to Gen AI evals, RAG Engine and Vector Search retained (accessed 2026-09-05)
- cloud.google.com/products/gemini-enterprise-agent-p… — Live pricing verified 05/09/2026: Sessions and Memory Bank billing 'will commence on September 1st, 2026', Semantic Governance Policy at one Agent Compute vCPU-hour ($0.085) per 15,000 response evaluations and billing 'later in 2026', Explainable AI billed at inference rates (accessed 2026-09-05)
- cloud.google.com/blog/topics/google-cloud-next/welc… — Official Kurian keynote transcript 22/04/2026: 40% quarter-over-quarter growth in paid Gemini Enterprise MAUs, Agent Studio, Agent-to-Agent Orchestration, Agent Registry, Agent Identity, Agent Gateway and Agent Observability named as new components; Gemini 3.1 Pro, Nano Banana 2, Lyria 3 and Claude Opus 4.7 on the platform (accessed 2026-09-05)
- cloud.google.com/blog/products/compute/cross-cloud-… — Official: Agent Gateway 'natively understands agent protocols like MCP and A2A to inspect and govern every agent interaction', integrating with Model Armor and Context Aware Access (accessed 2026-09-05)
- docs.cloud.google.com/gemini-enterprise-agent-platf… — Official Gen AI evaluation service docs: rubric-based quality metrics, sampling directly from production logs, and synthetic data generation for eval sets (accessed 2026-09-05)
- docs.cloud.google.com/gemini-enterprise-agent-platf… — Official REST reference confirming generateSyntheticData exists as a documented API under Gen AI evaluation, scoped to task or data descriptions (accessed 2026-09-05)
- google.github.io/adk-docs — Official ADK docs: open source, Python / TypeScript / Go / Java / Kotlin, graph workflows 'new in ADK 2.0', Agents CLI, structured context management, one-command deploy to Agent Runtime, Cloud Run or GKE with Cloud Trace observability inherited (accessed 2026-09-05)
- cloud.google.com/blog/products/ai-machine-learning/… — Official SKU change table: runtime $0.0994 to $0.0864 per vCPU-hour from 16/12/2025, Code Execution free to $0.0864 and Sessions free to $0.25 per 1,000 events from 28/01/2026, Memory Bank retrieval $0.50 per 1,000; Cloud API Registry as a private tool registry (accessed 2026-09-05)
- docs.cloud.google.com/gemini-enterprise-agent-platf… — Official overview describing Agent Platform as the evolution of Vertex AI with over 200 foundation models across the full AI lifecycle (accessed 2026-09-05)
- cloudzero.com/blog/google-vertex-ai-pricing — Independent cost analysis: the four additive SKUs behind one user interaction (search $4-6 per 1,000, model tokens, sessions $0.25 per 1,000 events, runtime $0.0864 per vCPU-hour), free tier of 50 vCPU-hours and 100 GB-hours (accessed 2026-09-05)
- betterclaw.io/blog/google-vertex-ai-agent-builder — Independent pricing breakdown updated 19/06/2026: realistic $500-$2,000+/month for a production support agent, reported surprise invoices into five figures, aggregate user rating ~4.3/5 with pricing complexity and learning curve the two most common complaints (accessed 2026-09-05)
- workos.com/blog/google-vertex-ai-vs-workos-ml-platf… — Independent analysis of the identity layer: Agent Engine agent identity is preview, 'Agentic IAM' is announced not GA, and the identity model governs Google Cloud resources rather than external IdPs or non-GCP MCP servers (accessed 2026-09-05)
- itpro.com/technology/artificial-intelligence/google… — Independent trade coverage of the consolidation: Agent Platform subsumes Vertex AI features and adds DevOps, orchestration and security, with low-code Agent Studio, open-source ADK and 200+ models (accessed 2026-09-05)
- gcpstudyhub.com/blog/vertex-ai-replaced-by-gemini-e… — Independent practitioner note 02/05/2026 on the rename's practical edges: the API is still aiplatform.googleapis.com, and classical ML products now sit as a sub-section of a platform named for agents (accessed 2026-09-05)
- engini.ai/blog/vertex-ai-agent-builder-guide-2026 — Independent guide 30/04/2026 listing limitations: narrower regional availability than open-source alternatives, billing unpredictability across vCPU, memory, search queries and model tokens, custom integrations needing engineering effort (accessed 2026-09-05)