WorkflowStack AI
WorkflowsIndustriesToolsGuidesAI QuizBlogEnterprise
Get Free Workflows
WorkflowStack AI

Practical AI workflows for SMB operators and enterprise teams. No fluff. No hype. Just what ships.

Library

  • All Workflows
  • Industries
  • Enterprise
  • Tools
  • Guides

Company

  • About
  • Blog
  • Newsletter
  • Contact

Stay Updated

Weekly workflow ideas for operators and enterprise teams.

Get Free Workflows

© 2026 Blueteem LLC. All rights reserved.

Privacy PolicyTerms of Service
HomeBlogBedrock AgentCore vs Databricks Mosaic AI vs Google Agent Platform: 2026 Pricing
November 10, 2026

Bedrock AgentCore vs Databricks Mosaic AI vs Google Agent Platform: 2026 Pricing

Where to build, host and govern production agents on AWS, Databricks or Google Cloud, with every published per-unit rate and a worked customer-service agent bill.

The short answer

At 100,000 customer-service sessions a month, the model costs more than the platform. Platform charges are up to $493 on Amazon Bedrock AgentCore, and about $319 on Google's Gemini Enterprise Agent Platform (formerly Vertex AI) at 1 vCPU and 1 GiB, or $1,253 at Google's 4 vCPU default. Tokens cost about $6,000. Databricks Mosaic AI (now Agent Bricks) publishes no Agent Runtime price; if it is billed like a default Databricks App, it would be about $270 per always-on instance.

What each platform is now called

Google. Google's release notes for 22 April 2026 ("Vertex AI to Gemini Enterprise Agent Platform naming changes") map "Vertex AI Platform" to "Agent Platform" and "Vertex AI Agent Engine" to "Agent Runtime". The old vertex-ai/pricing URL redirects to the Agent Platform pricing page. Tool page: Google Vertex AI.

Databricks. Databricks' agent documentation (updated 13 October 2026) calls Agent Bricks "the Databricks agent developer platform"; the "Mosaic AI" name does not appear on that page or the pricing pages read for this post. Tool page: Databricks Mosaic AI. The Agent Bricks Knowledge Assistant and Supervisor Agent "are no longer available to new customers"; Databricks recommends Genie Agents instead.

AWS. Amazon Bedrock AgentCore keeps its name: separately metered Runtime, Gateway, Memory, Identity, Observability, Browser, Code Interpreter, Policy, Evaluations, Web Search and others.

All three sit in our AI infrastructure category.

Pricing at a glance

Prices and plan names are as published on each vendor's own pricing page; last verified 10 November 2026. Databricks rates are Premium plan, AWS, US East (the page default).

Pricing at a glance
VendorBilling unitPublished entry priceWhat it's for
AWS Bedrock AgentCore Runtime (v2)vCPU-hour and GB-hour, per second, active CPU only$0.1276 per vCPU-hour; $0.0169 per GB-hourServerless, session-isolated agent hosting on AWS
Google Agent RuntimeAgent Compute vCPU-hour and Agent Memory GiB-hour, on allocated resources$0.085 per vCPU-hour; $0.009 per GiB-hour (deploy-doc default container: 4 vCPU, 4 GiB)Agent hosting on Google Cloud, Gemini-first
Databricks Agent Runtime (runs on Databricks Apps)No Agent Runtime price published; Apps bill App Capacity DBUsInferred: default Medium app 0.5 DBU/hr × $0.75 = $0.375/hr (confirm with Databricks)Agents next to Unity Catalog data
Databricks Model Serving (CPU)DBU, $0.07 per DBUStandard workloads billed at 4 DBU per hourServing models, features or agents as endpoints
Microsoft Foundry Agent ServiceHosted agents: vCPU-hour and GiB-hour$0.0994 per vCPU-hour; $0.0118 per GiB-hour (East US 2)Agents on Azure; prompt agents carry no extra charge

AgentCore: every component metered separately

Runtime. v2 costs $0.1276 per vCPU-hour and $0.0169 per GB-hour, billed per second with a 128 MB memory floor. CPU "scales to zero during I/O wait", but a session is billed "until session termination", and v2 reclaims idle memory after 120 seconds. Committed-baseline rates ($0.0997 and $0.0132) are still described as "launching by October 2026".

Gateway and Identity. Gateway is $0.005 per 1,000 invocations, $0.025 per 1,000 searches and $0.02 per 100 tools indexed a month. Identity is free through Runtime or Gateway.

Memory. Short-term: $1.00 per GB ingested, $0.20 per GB retrieved, $0.10 per GB-month stored, with each event billed at a 12 KB minimum. Long-term: $0.75 per 1,000 records a month (built-in extraction) or $0.25 (self-managed), plus $0.50 per 1,000 retrievals.

Everything else. Browser and Code Interpreter are $0.0895 per vCPU-hour; Policy is $0.000025 per authorization request plus $0.13 per 1,000 input tokens for policy authoring; Observability is billed at CloudWatch rates. Model inference is separate, through Amazon Bedrock or another provider.

Google Agent Platform: one compute meter for everything

Google prices its agent services through Agent Compute ($0.085 per vCPU-hour), Agent Memory ($0.009 per GiB-hour) and Agent Storage ($0.30 per GiB-month). 1- and 3-year Flexible Savings Plans lower compute to $0.0765 and $0.068.

Runtime is billed on allocated vCPU and memory, rounded to the second, and "idle time spent waiting for the next prompt between turns is not billed." Allocation is per container: Google's deploy documentation sets a default of 4 vCPU and 4 GiB per container, serving up to 9 requests at once, with min_instances defaulting to 1. Idle minimum instances are not billed while that feature is in Preview, which Google says "is subject to change".

Other services convert requests into compute hours: 15,000 Agent Gateway calls, or 1 million Sessions or Memory Bank writes, or 3 million reads, each equal one vCPU-hour. Memory Bank's memory-generation and embedding tokens are "billed separately under their respective model SKUs". Sessions and Memory Bank billing began on 1 September 2026; Semantic Governance Policy billing starts "later in 2026".

Get one new AI workflow every week

Practical playbooks, prompts, and tool stacks — for the SMB operators and enterprise teams actually shipping AI.

Databricks: DBUs, and hosting that runs by the hour

Databricks' Agent Runtime runs your agent "on Databricks-hosted serverless compute" via Databricks Apps, on 1 to 5 instances. Databricks publishes no Agent Runtime price. Apps bill App Capacity at $0.75 per DBU, and the default Medium instance (up to 2 vCPUs, 6 GB) uses 0.5 DBU an hour. If Agent Runtime is billed the same way, which you should confirm with Databricks, that is an inferred $0.375 an hour, or $270 a month, per always-on instance.

The alternative, a Model Serving CPU endpoint, is billed at "4 DBU per hour" for standard workloads ($0.28 an hour at $0.07 per DBU). The page's 30-minute scale-to-zero timeout is stated for GPU Serving, so confirm CPU behaviour with Databricks.

Agent memory is stored in Lakebase: $0.069 per capacity-unit hour as the always-on minimum, $0.092 for autoscaling usage (50% off until 31 January 2027), and $0.345 per GB-month of storage. Unity Gateway charges only for Usage Tracking and Inference Tables ($0.10 per GB each), with "no additional Unity Gateway charges to manage assets such as models or MCPs". GPT-6.1 Sol short-context costs 28.571 DBU per million input tokens and 142.857 per million output: $2.00 and $10.00 at $0.07 per DBU, the same as OpenAI's list price.

Microsoft Foundry Agent Service, briefly

Prompt and workflow agents carry "no additional charge"; you pay for model tokens and tools. In East US, Code Interpreter is $0.033 per session, Web Search $14 per 1,000 transactions and File Search storage $0.11 per GB per day (1 GB free). Hosted agents built on frameworks such as LangGraph cost $0.0994 per vCPU-hour and $0.0118 per GiB-hour in East US 2; the page shows N/A for Central US, its default region.

Worked example: one customer-service agent

Assumptions:

  • 100,000 six-minute sessions a month
  • 120 seconds of turn processing per session (including model and tool waits), 36 seconds of it active CPU; the other 4 minutes are customer think time
  • AgentCore: 1 vCPU and 1 GB of memory per session
  • Google: resource_limits set to 1 vCPU and 1 GiB (default 4 vCPU, 4 GiB), billed for the 120 seconds of turns
  • 8 tool calls and 1 tool search per session
  • 10 short-term memory events of 3 KB each per session, each read back once
  • 1 long-term memory record stored and 2 retrieved per session
  • short-term events kept 30 days; long-term records expire after a month (else storage grows $75 a month)
  • 20,000 input tokens and 2,000 output tokens per session, with no prompt caching
  • sessions end when the conversation ends; CloudWatch, Cloud Logging and data transfer excluded

AWS AgentCore (v2 consumption):

  • CPU: 36 s × $0.1276 ÷ 3,600 = $0.001276 per session, so $127.60 a month
  • Memory (upper bound): 360 s × 1 GB × $0.0169 ÷ 3,600 = $0.00169 per session, so up to $169.00 a month; v2 reclaims memory idle for over 120 seconds, down to a 128 MB floor, so the real figure can be lower
  • Gateway: 800,000 calls × $0.005 per 1,000 = $4.00, plus 100,000 searches × $0.025 per 1,000 = $2.50
  • Short-term memory: 1,000,000 events at the 12 KB minimum = 12 GB; ingestion $12.00, retrieval $2.40, storage 3 GB × $0.10 = $0.30
  • Long-term memory: 100,000 records × $0.75 per 1,000 = $75.00, plus 200,000 retrievals × $0.50 per 1,000 = $100.00
  • Identity: $0 through Runtime or Gateway
  • Total: up to about $492.80 a month

Google Agent Runtime:

  • Compute: 120 s × $0.085 ÷ 3,600 × 100,000 = $283.33
  • Memory: 120 s × 1 GiB × $0.009 ÷ 3,600 × 100,000 = $30.00 (the between-turn wait is not billed)
  • Gateway: 800,000 ÷ 15,000 = 53.3 vCPU-hours × $0.085 = $4.53
  • Sessions: $0.085 for 1 million writes, $0.03 for 1 million reads, about $0.84 storage
  • Memory Bank: under $0.10 for operations and storage, plus memory-generation tokens (billed separately, not included)
  • Total: about $319 a month, plus memory-generation tokens

At Google's default 4 vCPU and 4 GiB container, the same arithmetic gives about $1,253 for compute and memory ($1,133.33 + $120.00); containers serve up to 9 requests at once, so real cost depends on concurrency.

Databricks (inferred): if Agent Runtime is billed like a default Medium Databricks App, one always-on instance costs 0.5 DBU × $0.75 × 720 hours = $270.00. One Lakebase capacity unit kept always on costs $0.069 × 720 = $49.68 at list price (autoscaling usage above the minimum is $0.092 per CU-hour), before the 50% promotion. Unity Gateway is $0 unless tracking is on. That totals about $320, but three inputs are unpublished:

  • whether Agent Runtime is billed as App Capacity
  • how many concurrent sessions one instance handles
  • how many Lakebase capacity units the memory store needs (managed agent memory is in Beta; during the preview only the Lakebase instance is billed, and pricing is "subject to change")

Model tokens (GPT-6.1 Sol short-context, OpenAI list price or the Databricks DBU equivalent; no caching): 2 billion input × $2 per million = $4,000, plus 200 million output × $10 per million = $2,000. That is $6,000 a month, roughly 12 to 19 times the platform bills above at 1 vCPU sizing.

When not to use each one

AgentCore. Look elsewhere if sessions sit open waiting on humans and you cannot end them promptly: sessions are billed until termination, and although v2 reclaims idle memory after 120 seconds, the 128 MB floor still accrues. Small events are billed as 12 KB, so 3 KB events cost four times their size. Long-term memory is the largest memory cost above: $100 of retrievals and $75 of built-in storage.

Google Agent Platform. Avoid it if you need predictable memory costs: Memory Bank's generation and embedding tokens are billed outside the Agent Compute meter. Runtime bills allocated vCPU during turns, so the 4 vCPU default costs four times the 1 vCPU sizing above unless you change resource_limits or rely on container concurrency. min_instances defaults to 1, and the Preview term that leaves idle minimum instances unbilled may change.

Databricks. It is a poor fit for spiky, low-volume agents on an always-on instance (an inferred $270 a month, paid whether or not sessions arrive), or for buyers who need a documented per-session price: Databricks publishes none for Agent Runtime.

Microsoft Foundry. Check region availability first: hosted-agent compute shows N/A in Central US.

Comparing model layers? See Amazon Bedrock vs Google Vertex AI and OpenRouter vs Bedrock vs Vertex AI.

Frequently asked questions

How much does Amazon Bedrock AgentCore cost?

AgentCore Runtime v2 is $0.1276 per vCPU-hour and $0.0169 per GB-hour, billed per second, with CPU charged only while active. Gateway is $0.005 per 1,000 invocations and long-term memory $0.75 per 1,000 records a month plus $0.50 per 1,000 retrievals; model inference is extra.

Is Vertex AI Agent Engine now called Agent Runtime?

Yes. Google's release notes for 22 April 2026 rename Vertex AI to Gemini Enterprise Agent Platform and Vertex AI Agent Engine to Agent Runtime, billed at $0.085 per vCPU-hour and $0.009 per GiB-hour of allocated resources.

Does Databricks still sell Mosaic AI?

Databricks now presents its agent tooling as Agent Bricks; the Mosaic AI name does not appear on the agent documentation or pricing pages read for this post. Agents deploy to Agent Runtime, which runs on Databricks Apps; Databricks publishes no Agent Runtime price, but Apps bill $0.75 per DBU and a default Medium app uses 0.5 DBU an hour.

Is Bedrock AgentCore cheaper than Google Agent Runtime?

It depends on the workload. In a 100,000-session customer-service example, platform charges were up to about $493 a month on AgentCore and about $319 on Google when containers are sized at 1 vCPU and 1 GiB (Google's default is 4 vCPU and 4 GiB, about $1,253). Model tokens, about $6,000 at GPT-6.1 Sol short-context rates, outweighed every platform bill.

Does Microsoft Foundry Agent Service charge for agents?

Microsoft says there is no additional charge for creating or running Foundry-native prompt and workflow agents; you pay for model tokens and tools. Hosted agents are billed at $0.0994 per vCPU-hour and $0.0118 per GiB-hour in East US 2.

Next step

For the full architecture, start with the multi-agent support deflection workflow. Then score these platforms with the Enterprise AI Agent Platform Buyer's Guide.

Related Workflows

Internal Developer Productivity Agent Platform

A governed internal platform for shipping coding and ops agents to engineering teams — with shared guardrails, evals, and observability instead of shadow tools.

View workflow

AI Agent Inventory & Governance Program

Find every agent your business units already shipped, right-size its permissions, and put a control plane around the ones that stay.

View workflow

Review Request Automation

Automatically ask happy customers for Google reviews at the perfect moment — right after the job.

View workflow

Keep Reading

November 10, 2026

Decagon vs Sierra vs Fin (Intercom) vs Parloa: Enterprise AI Customer Agents in 2026

Only Fin publishes a per-unit price ($0.99 per outcome). Decagon, Sierra and Parloa are quote-only, so this comparison focuses on what each one bills for and the break-even rates to bring to the negotiation.

November 10, 2026

Composio vs Arcade vs MCP in 2026: Agent Tool Integration Pricing, Auth and Security

The managed tool layers for AI agents price the same action four different ways. Here is what Composio, Arcade.dev and Pipedream Connect publish, what running your own MCP servers costs you instead, and a worked monthly bill.

November 10, 2026

OpenRouter vs Amazon Bedrock vs Google Vertex AI (2026): Router, Cloud, or Gateway?

OpenRouter is a hosted router, Bedrock and Vertex AI are cloud model platforms, and LiteLLM is a gateway you run yourself. Here is what each costs on the same Claude workload, and which one your cloud contract probably picks for you.

Found this helpful?

Get weekly AI workflow ideas in your inbox.