Your team has a working agent prototype and a decision to make by the end of the quarter: which cloud do you build it on for real? The shortlist is Amazon Bedrock AgentCore, Microsoft Foundry Agent Service and Google Vertex AI Agent Engine. Three sales teams have quoted you three numbers that are not comparable, and the finance director wants one page.
This is that page. It compares what each platform charges to run an agent, what the tools cost on top, and where each one loses.
Reviewed by the UD enterprise AI team. Prices are US list rates published by each vendor and checked on 4 September 2026. Last updated 4 September 2026.
What are you actually choosing between?
You are choosing an agent runtime, not a model. All three platforms let you run several model families, so the decision is about where the agent process executes, how memory and tools are billed, and which discount structure your finance team can commit to.
This is a different question from which assistant to license per seat. If your decision is about Copilot or Gemini licences for staff, that is a separate comparison.
It is also different from raw model pricing. The per-token cost of the flagship models is covered in our model cost comparison, and those charges sit on top of everything below.
The confusion in most vendor quotes comes from mixing these three layers. Separate them and the comparison becomes tractable.
How much does each platform charge to run an agent?
Two of the three meter compute by the vCPU-hour and gigabyte-hour, at rates within four per cent of each other. The third charges nothing for the orchestration layer itself. That single structural difference explains most of the gap between the quotes you have been given.
Agent runtime list prices, September 2026
--- Amazon Bedrock AgentCore. US$0.0895 per vCPU-hour and US$0.00945 per GB-hour, per AWS published pricing. CPU is billed for active cycles only; time spent waiting on a model or tool call is not billed as CPU.
--- Google Vertex AI Agent Engine. US$0.0864 per vCPU-hour and US$0.0090 per GB-hour, per Google Cloud published pricing. Marginally cheaper than AgentCore on both axes.
--- Microsoft Foundry Agent Service. No additional charge for the agent service itself, per Azure published pricing. You pay for model tokens and for every tool the agent invokes.
The practical reading is that AWS and Google are within rounding distance of each other on runtime, so runtime price should not decide this. Microsoft's zero platform fee is real, but it moves the cost rather than removing it.
What do the tools cost on top of the runtime?
Tool invocations are where quotes diverge sharply, because an agent that searches, remembers and calls APIs triggers metered events on every turn. Search is the single largest variable, and the published rates differ by more than four times between platforms.
Metered tool rates, September 2026
--- Web or enterprise search. AgentCore web search is US$7.00 per 1,000 queries. Vertex AI Search runs US$1.50 to US$6.00 per 1,000 queries depending on tier. Foundry bills grounding and knowledge connections such as Azure AI Search and Bing grounding separately, priced per connected service.
--- Memory events. Both AgentCore short-term memory and Vertex Agent Engine session and memory storage are US$0.25 per 1,000 events.
--- Gateway and API invocation. AgentCore gateway invocation is US$0.005 per 1,000 calls, which is close to a rounding error at most enterprise volumes.
Memory pricing is identical across AWS and Google, which is a useful signal. When two competitors land on the same number, that component is commoditised and should not appear in your evaluation matrix at all.
Search should. An agent handling 200,000 grounded queries a month costs about US$1,400 on AgentCore web search and as little as US$300 on the lowest Vertex AI Search tier, before any other line item.
How do the discount structures compare?
All three sell consumption with no upfront commitment, then discount heavily for committed volume. The differences are in commitment length and flexibility, and at enterprise scale this layer usually matters more than list price.
Google Vertex AI offers committed use discounts reported at up to 57 per cent, plus context caching for Gemini models at a 75 per cent discount on cached input tokens. For a retrieval-heavy agent that re-reads the same documents, caching is the largest single lever available.
Azure offers provisioned throughput and reserved capacity, with reported savings of 30 to 50 per cent for predictable workloads above roughly one million tokens a day. Azure's commit structures are generally the most flexible of the three.
AWS pairs pay-per-use with the deepest reserved-instance ecosystem, which matters if your agent workload sits alongside substantial existing AWS compute you already commit to.
The honest summary is that discount eligibility usually depends on total cloud spend, not agent spend. Your negotiating position is set by the rest of your estate.
Which platform should you choose?
The correct answer for most Hong Kong enterprises is determined before the pricing conversation starts, by where your data and identity already live. Migration cost and integration effort exceed the runtime price gap by an order of magnitude.
Verdict by buyer type
--- Already standardised on Microsoft 365 and Entra ID. Choose Foundry. The zero platform fee is secondary; the real saving is that identity, permissions and SharePoint grounding already work.
--- Retrieval-heavy agents over large document sets. Choose Vertex AI. The combination of lower search tiers and 75 per cent cached-input discounting compounds fastest as document volume grows.
--- Existing significant AWS estate with data in S3. Choose Bedrock AgentCore. Keeping the agent next to the data avoids egress charges that will dwarf the runtime difference.
--- Regulated industry needing customer-held monitoring data. Any of the three. Anthropic's Enterprise Frontier Safeguards will be supported on Bedrock, Microsoft Foundry and Google's Agent Platform, so this requirement no longer narrows the field.
--- No strong existing cloud commitment. Choose Vertex AI on price, but treat this as the rarest case. Enterprises with 50 to 500 staff almost always have a commitment already.
Where does each platform lose?
Every platform in this comparison has a real weakness, and a vendor who cannot name theirs is not being straight with you. These are the limitations most likely to surface after the contract is signed rather than before.
Named limitations
--- Bedrock AgentCore. The highest published search rate of the three at US$7.00 per 1,000 queries, and the most components to price separately, which makes forecasting harder for teams without FinOps support.
--- Microsoft Foundry. The zero platform fee is genuinely zero, but every meaningful tool is a separately billed Azure service. Teams routinely underestimate the total because the headline number is attractive.
--- Google Vertex AI. The lowest runtime rates, but the deepest discounts require committed use terms that a finance director may not sign in year one of an unproven workload.
--- All three. None of these prices include model tokens, which are usually the largest line on the invoice. A runtime comparison that ignores tokens is not a cost comparison.
One further caveat applies to everything above. These are US list rates published by each vendor. They change, they vary by region, and Hong Kong deployments should be priced against the specific regional endpoint you intend to use.
What does a realistic monthly workload cost?
Work an example rather than comparing rate cards. Take a mid-sized Hong Kong logistics firm running one internal agent that handles shipment queries: roughly 500 vCPU-hours and 1,000 GB-hours of active compute a month, 200,000 grounded searches and 400,000 memory events.
Worked example, runtime and tools only, excluding model tokens
--- Bedrock AgentCore. About US$45 compute, US$9 memory, US$1,400 web search. Roughly US$1,455 a month, dominated almost entirely by search.
--- Vertex AI Agent Engine. About US$43 compute, US$9 memory, US$300 to US$1,200 search depending on tier. Roughly US$350 to US$1,250 a month.
--- Microsoft Foundry. No platform fee, so the bill is your Azure AI Search instance plus grounding services, which is why a Foundry quote is impossible to compare without knowing which services you attach.
Three observations follow. Compute is trivial at this scale. Memory is identical between AWS and Google. Search is the entire decision, and it is the line item vendors are least likely to lead with.
Now add model tokens. At typical enterprise volumes the token bill for this workload runs several times the runtime and tool total combined, which is why a platform comparison conducted without token modelling produces the wrong answer with great confidence.
Does the platform choice actually decide the outcome?
Usually not. Gartner's 2026 CIO and Technology Executive Survey found only 17 per cent of organisations have deployed AI agents to date, while more than 60 per cent expect to within two years. The gap is not caused by pricing pages.
S&P Global Market Intelligence data cited alongside Gartner's work puts 31 per cent of organisations with an agent running in production, against Gartner's finding that 80 per cent of enterprise applications shipped or updated in the first quarter of 2026 embed at least one agent.
Read those two numbers together and the pattern is clear. Agents are being shipped into products far faster than organisations are putting them into production, and the constraint is governance, integration and adoption rather than the cost per vCPU-hour.
Gartner's 2026 Hype Cycle for Agentic AI reinforces this, placing agentic AI governance, agentic AI security and FinOps for agentic AI as rising concerns across the cycle.
The implication for your decision is uncomfortable but useful. Optimising the runtime rate saves single-digit percentages. Choosing a platform your team cannot operate safely costs the entire programme.
What is the correct next step?
Do not start with a pricing spreadsheet. Start by establishing which platform your existing identity, data residency and security posture already point to, then price only that option properly against one realistic workload.
--- Step one. Write down where your production data sits today and which identity provider governs access. This eliminates at least one platform immediately in most organisations.
--- Step two. Model one real workload end to end, including model tokens, search queries, memory events and egress, rather than comparing headline rates.
--- Step three. Ask each shortlisted vendor for the regional rate for your Hong Kong endpoint in writing, and for the commitment threshold that unlocks their discount tier.
--- Step four. Before committing, confirm your organisation can actually operate an agent in production: monitoring, escalation, change control and an owner with a name.
Step four is where most programmes fail, and it is the one step no vendor pricing page will prompt you to take.
The bottom line
On published rates, Vertex AI Agent Engine is marginally the cheapest runtime, Foundry has no platform fee at all, and Bedrock AgentCore carries the highest search rate. The spread between them is small enough that runtime price should be the last input to your decision, not the first.
What should decide it is where your data already lives, which identity system already governs it, and whether your team can operate what you build. Those three answers are usually the same answer, and they are rarely the one in the cheapest quote.
We understand AI. We understand you. With UD by your side, AI never feels cold. Twenty-eight years alongside Hong Kong enterprises has taught us that the platform is the easy part.
Before you commit to a platform
A free AI readiness assessment tells you whether your organisation can operate an agent in production before you sign a committed-use agreement. We'll walk you through every step, from readiness assessment and workload modelling to platform selection, deployment and performance tracking, with twenty-eight years of Hong Kong enterprise experience behind you.
