Google Vertex AI Agent Builder has no flat platform fee. As of September 2026, you pay for what your agents consume: compute on Agent Engine bills at $0.0864 per vCPU-hour and $0.0090 per GB-hour of memory, sessions and memory events cost $0.25 per 1,000, and the model calls your agents make bill separately at standard Gemini rates. New Google Cloud customers get $300 in credit, and Express Mode lets you start building without a billing account at all.
The pricing has three layers, and it helps to keep them separate. The first layer is compute. When you deploy an agent to Agent Engine, Google charges for the vCPU and memory the runtime consumes, not per seat or per agent. An agent holding one vCPU and 2 GB of memory around the clock costs about $63 a month for the compute and $13 for the memory, roughly $76 before the agent does anything intelligent.
The second layer is state. Sessions and Memory Bank, the pieces that let an agent remember a user across conversations, bill at $0.25 per 1,000 sessions or memory events. At most usage levels this is a rounding error, but it is the line item that scales directly with how many conversations you run rather than how big your deployment is.
The third layer is the one that dominates in production: model usage. Every reasoning step, tool call, and response your agent generates consumes Gemini tokens, billed at the standard per-token rates for whichever model you pick. In my testing, token spend outweighs compute spend quickly once an agent handles real traffic, so any serious cost estimate starts with expected conversation volume and tokens per turn, not with the infrastructure math.
Two on-ramps soften the entry cost. The $300 new-customer credit applies to Agent Builder like any other Google Cloud service, which covers a meaningful amount of prototyping. Express Mode goes further: you can sign up and build within free-tier quotas without attaching a billing account, then upgrade when you outgrow it. That makes the realistic evaluation cost close to zero.
The structure tells you who Google built this for. There is no per-conversation price and no per-agent license, which makes costs harder to forecast than seat-based tools but cheaper at scale for teams that already think in cloud units. The question is not whether the rates are fair. It is whether you can predict your agents' token appetite before you commit.
Go deeper: my full Vertex AI Agent Builder review and why token costs decide AI TCO.