Gemini Enterprise Agent Platform

Positioning: one product, two eras of branding

Google Cloud's AI platform is in the middle of a rebrand that will confuse buyers for another year. What this directory lists as Vertex AI now goes to market as the Gemini Enterprise Agent Platform — the rename happened around Google Cloud Next 2026, where Google folded Vertex AI, Agentspace and Gemini Code Assist Enterprise into a single control plane for building, governing and deploying agents. Google's own language is that this is a rebrand rather than a retirement: existing Vertex AI endpoints, notebooks, training and prediction surfaces continue, now documented under the new product name. The page I read this week sells it as a platform to "innovate, build, and deploy enterprise-ready agents and apps", with Google's Antigravity agent orchestration app available to steer multiple agents through whole workflows from the same Google Cloud credentials.

Under the new name, this remains the broadest single-vendor ML platform I track: frontier Gemini models plus a model garden of 200-plus third-party models, agent tooling, AutoML, custom training and prediction, notebooks, and governance — all metered on the Google Cloud invoice.

What it actually does

The generative AI layer is the front door. Gemini 4 is the current headline model, priced at $2 per 1M input tokens and $10 per 1M text output on global processing at time of writing (non-global is $2.20/$11.00), with a promotional 50% credits-back running through 30 November 2026; Gemini 4 Argon follows on 1 December 2026 at $4/$20 global. Cached input is $0.10 per 1M. The model garden reaches well beyond Gemini: Llama 4 Scout at $0.25/$0.70, Llama 4 Maverick at $0.35/$1.15, Llama 3.3 70B at $0.72/$0.72, OpenAI's open models gpt-oss-120b at $0.09/$0.36 and gpt-oss-20b at $0.07/$0.25, and xAI's Grok line from Grok 4.1 Fast at $0.20/$0.50 up to Grok 4.7 at $2.00/$6.00 under 200k tokens. Batch rates run at half the standard token price for supported models.

The agent layer is where the platform is pushing hardest. Agent Runtime executes agents, Agent Compute meters that execution at $0.085 per vCPU-hour, and one vCPU-hour also covers 15,000 Agent Gateway API or authorisation calls. Memory Bank charges $0.085 per 1M write operations for long-term memory maintenance, with model tokens for memory generation billed separately. Semantic Governance Policy — Google's layer for evaluating agent responses against policy — meters at $0.085 per 15,000 evaluations plus the evaluation model's tokens. The Gemini Deep Research Agent is listed at $2 input / $12 output per 1M tokens. Those numbers come from Google's generative AI pricing page.

Then the classic ML platform underneath: AutoML for tabular, image, video and text data (tabular forecasting training is listed at $21.252 per node hour on the 1-year savings plan rate), supervised fine-tuning for Gemini and open models (Gemini 3.5 Flash Lite fine-tuning at $0.003 per 1,000 training tokens; Llama 4 Scout fine-tuning at $5.77 per 1M), custom-trained model deployment on managed endpoints with 30-second billing increments and no minimum duration, notebooks via Colab Enterprise or Workbench with native BigQuery integration, and Model Garden for discovery. New customers get $300 in free credits across the platform. Storage, networking and compute bills are the usual Google Cloud lines underneath all of it.

My testing notes

I have used the platform across its two names for evaluation and deployment work. Four observations.

First, the agent metering model takes real attention. The unit prices are tiny — $0.085 per vCPU-hour is nothing in isolation — but the meters multiply: Agent Compute, Memory Bank writes, governance evaluations, and then the model tokens underneath each of those. In a test deployment of a memory-using agent, my cost decomposed roughly into a fifth platform meters and four fifths model tokens. That is a reasonable split, but it only became visible after I instrumented the spend per meter; the billing console shows the lines, not the causal chain.

Second, the model garden pricing is competitive and the breadth is real. Routing the same summarisation prompt across Llama 4 Scout at $0.25/$0.70 and gpt-oss-20b at $0.07/$0.25 on the same endpoint interface was a config change, and the gpt-oss run cost a fraction of the Scout run for output of comparable usefulness on that task. For teams that treat models as interchangeable commodities, this is the strongest argument for the platform. The trade-off is version lag: the garden is a distribution channel, and channels move at distribution speed.

Third, the rebrand is not cosmetic in the documentation. Several deep links from old guides resolve to redirected pages, and one AutoML pricing table I read carried the new product name over legacy Vertex labels with consumption model IDs attached. I would not migrate anything on the assumption that today's URLs survive; pin your infrastructure-as-code to API versions rather than console paths.

Fourth, latency and reliability on Gemini endpoints matched expectations for a hyperscaler in my runs: my parallel request test held steady at moderate concurrency, and I did not see the queueing behaviour I have measured on some single-model providers. The 50%-credits-back promotion on Gemini 4 through 30 November 2026 is worth planning around — I confirmed the promotional framing on the pricing page, and it materially changes near-term unit economics for teams deciding between Gemini 4 today and Argon pricing in December.

Fifth, the governance story deserves a practical note. I configured a Semantic Governance Policy against a customer-facing agent and it scored a sample of responses against the written policy at $0.085 per 15,000 evaluations plus evaluation-model tokens — trivially cheap at my sample sizes, and genuinely useful as a guardrail before responses reach users. The limitation is that it evaluates what you tell it to evaluate; a policy that omits a failure mode will happily pass that failure mode through. I treat it as an automated reviewer with a narrow brief, not as compliance evidence on its own.

What it does not do

There is no flat-rate seat pricing; everything is consumption, and the number of metered surfaces means a naive deployment can produce an invoice nobody predicted. The platform assumes Google Cloud competence — IAM, project structure, service accounts and quota management are prerequisites, not features. AutoML and custom training remain comparatively expensive for small experiments (tabular forecasting training alone is over $20 per node hour on the savings rate). Governance tooling is powerful but metered and configured per policy; the Semantic Governance layer is only as good as the policy definitions you write. Model availability varies by region, and some third-party garden models lag the vendors' own APIs on version freshness — Grok 4.7 arrived here after xAI's own endpoint. And if your team is not already a Google Cloud shop, the identity and networking work to get here safely is its own project.

Verdict

For ML engineering teams building agents and models on Google Cloud, the renamed platform does more under one invoice than anything else I evaluate: frontier models, a priced model garden, agent runtime with memory and governance, AutoML and custom training, and notebooks wired to BigQuery. It is consumption-priced, meter-heavy and unapologetically complex, and I would not recommend it to a team that is not already living in Google Cloud. I rate it 4.1: core functionality is deep across both agent and classic ML workloads, value is good at the garden's low end and depends on meter discipline at the agent layer, usability rewards cloud-native teams and punishes everyone else, the ecosystem and support apparatus is hyperscaler-scale, and reliability signals are strong — with the rebrand itself as the main short-term risk to anything you automate against it.

Sources

• https://cloud.google.com/products/gemini-enterprise-agent-platform
• https://cloud.google.com/products/gemini-enterprise-agent-platform/pricing
• https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing
• https://luby.co/blog/google-cloud-next-2026-what-the-gemini-enterprise-rebrand-means-for-you
• https://optimnow.io/post/gemini-pricing-explained-ai-studio-vs-vertex-ai
• https://betterclaw.io/blog/vertex-ai-agent-builder-review
• https://engini.ai/blog/vertex-ai-agent-builder-guide-2026
• https://markaicode.com/pricing/vertex-ai-pricing
• https://zarifautomates.com/blog/google-cloud-ai-for-enterprise-platform-overview
• https://antigravity.google/docs/enterprise