Short answer: Amazon Bedrock AgentCore can be inexpensive to start, but most real deployments cost more than the runtime line item suggests. AWS prices Runtime, Browser, and Code Interpreter by active CPU and memory usage, then separately charges for Gateway calls, memory, policy checks, evaluations, registry usage, and any wallet-provider fees for payments, while observability, storage, network transfer, and model inference can sit outside the core AgentCore rate card. For most buyers, the right budget question is not “What is the sticker price?” but “What will one completed workflow cost after tools, models, monitoring, and rollout work are included?”
That distinction matters even more now that AgentCore is becoming a deeper part of the AWS workflow stack. As of June 3, 2026, AWS Step Functions can add AgentCore-powered reasoning steps into workflow execution, which means Bedrock and AgentCore costs can show up inside broader automation budgets rather than in a standalone pilot.
November 15 update: Runtime V1 and V2 are different
Since November 12, new runtimes default to V2 in the regions listed in the AWS announcement, including US East (N. Virginia and Ohio). Existing runtimes stay on their current version until migrated. Explicit platformVersion selection controls V1 versus V2; check your region and deployed configuration rather than assuming every session changed.
V2 reclaims idle memory after 120 seconds. Higher unit prices therefore do not establish a higher total bill: compare measured billable usage for the same completed workload. Browser and Code Interpreter have separate rates and should not inherit V2 Runtime prices.
What the AWS rate card actually measures
AgentCore is modular. You can use one component or combine several. That is good for flexibility, but it also means there is no single all-in AgentCore price. Buyers need to model the specific services their agent will actually touch.
Amazon Bedrock AgentCore pricing components buyers should model
| Component | How pricing works | What changes the budget fastest |
|---|---|---|
| Runtime microVMs | V1: $0.0895/vCPU-hour and $0.00945/GB-hour. V2 consumption: $0.1276/vCPU-hour and $0.0169/GB-hour. Verify current regional billing. | Session length, true active compute time, peak memory, and concurrency |
| Browser | $0.0895/vCPU-hour and $0.00945/GB-hour, separately from Runtime V2 | Long browser sessions, web latency, and memory-heavy tasks |
| Code Interpreter | $0.0895/vCPU-hour and $0.00945/GB-hour, separately from Runtime V2 | Execution count, runtime duration, and memory usage |
| Gateway | $0.005 per 1,000 API invocations, $0.025 per 1,000 search calls, and $0.02 per 100 indexed tools per month | How often the agent discovers and calls tools |
| Memory | Short-term events, long-term storage, and retrieval are billed separately | How much context you persist and how often the agent recalls it |
| Evaluations and policy | Token-based and request-based pricing | How aggressively you test, monitor, and govern production agents |
The important takeaway is that AgentCore is infrastructure pricing, not seat pricing. If you need predictable per-user licensing and very little engineering ownership, the AWS bill can be only one part of the total cost story.
Where Bedrock AgentCore budgets expand faster than teams expect
Model inference is not the same thing as AgentCore
AgentCore helps run the agent loop and surrounding services, but you still need to pay for the model work that powers reasoning and generation. If your workflow is output-heavy, uses expensive models, or calls multiple models in a single task, model spend can overtake the AgentCore runtime layer quickly.
Tool use is cheap per call until the workflow scales
Gateway pricing looks small on paper, but tool-heavy agents multiply calls fast. A workflow that checks order status, updates tickets, writes CRM notes, queries policy, and requests approval may create several billable events for every customer interaction. That is still manageable at pilot volume, but it matters once traffic reaches hundreds of thousands or millions of sessions.
Observability, storage, and network costs sit nearby
CloudWatch pricing applies to AgentCore observability, and AWS separately notes storage and data transfer charges in several parts of the stack. That means the clean-looking AgentCore rate card is not the same thing as your full operating bill.
Implementation effort is outside the AWS bill
Even if the platform cost is reasonable, businesses still need workflow design, tool wiring, permissions, QA, guardrails, monitoring, and ownership after launch. AgentCore can improve infrastructure efficiency, but it does not remove the need to operationalize the agent.
Example budget scenarios buyers can model
These are planning scenarios, not quotes. They are useful because they show how fast cost shape changes once you move from a simple text-and-tool agent to a browser-heavy or code-executing workflow.
Illustrative compute-only worksheet: Assume your monthly meter records 100 vCPU-hours and 200 GB-hours. At V1 rates the arithmetic is (100 × $0.0895) + (200 × $0.00945) = $10.84. Holding those same billed quantities constant at V2 consumption rates gives (100 × $0.1276) + (200 × $0.0169) = $16.14. These are our deliberately simple planning inputs, not an AWS workload benchmark, a quote, or a migration forecast.
The equal-usage assumption isolates rate differences but does not estimate V2 memory savings. Export the actual billable CPU and memory for a representative pilot, then calculate both a typical case and a retry-heavy case. Keep model tokens, tools, storage, transfer, monitoring and engineering effort separate. Divide the complete cost by successfully completed workflows, not raw sessions.
Correction: Earlier versions quoted an older AWS Runtime example as $723.50 per million sessions without its full workload assumptions. That figure is not a current general budget estimate and should not be used to price V2. The worksheet above replaces it rather than presenting a different AWS example as an equivalent workload.
A useful planning habit is to build the budget in layers: platform services first, model inference second, monitoring and governance third, and internal implementation effort last. Many teams do the reverse and underestimate total cost.
A simple ROI and payback formula
Use this plain-language formula:
Monthly ROI = (monthly value created - monthly operating cost) divided by monthly operating cost.
Payback period in months = one-time setup cost divided by monthly net benefit.
Example: if an internal support agent saves 800 labor hours per month, and those hours are worth $25 each, that is $20,000 in monthly value. If the full monthly operating cost of AgentCore, Bedrock models, monitoring, and support is $6,000, then monthly net benefit is $14,000. If initial setup cost was $28,000, payback is about two months.
The mistake is using raw interaction volume as the success metric. ROI comes from completed work, labor removed, faster cycle times, higher containment, or revenue protected. If the agent touches many sessions but resolves little, even a modest usage bill can become a bad investment.
How to decide if Bedrock AgentCore is worth it
AgentCore is usually strongest when you want custom AWS-native agents, need flexible infrastructure instead of bundled seat pricing, and can measure a high-value workflow with clear savings. It is especially compelling when agents spend a lot of time waiting on I/O, because AWS charges Runtime on active resource consumption rather than simple pre-allocated capacity.
It is less attractive when your organization mainly wants a plug-and-play business tool with minimal engineering ownership, highly predictable per-user pricing, or a fast rollout for a narrow use case like customer support chat alone. In those cases, a packaged agent platform or a managed implementation can be easier to budget even if the raw infrastructure line item looks higher.
If you are comparing Bedrock AgentCore with a packaged AI agent product, the real decision is build economics versus operating simplicity. AgentCore can lower infrastructure waste, but buyers still need enough workflow value to justify the extra design and ownership responsibility.