What this AgentCore cost estimate covers
The total includes AgentCore Runtime, Memory, Gateway, Knowledge Base (optional), Observability (optional), and model inference for the foundation model you select. Bedrock Prompt Caching is not modeled, but can reduce input token costs by up to 90 percent for repeated context. CloudWatch charges for Observability route through your CloudWatch account and are estimated at 15 percent of AgentCore cost. Data transfer and ECR storage for container deployments are not included.
If you're using custom memory extraction strategies or overriding the default model for long-term memory consolidation, your Memory costs will be higher than shown. The default built-in strategies include model costs in the per-record price. based on the assumptions entered.
Your foundation model choice matters more than any other input. At typical session lengths, model inference accounts for the majority of the total. Switching from Claude Sonnet to a smaller model can cut that line item in half. The calculator lets you compare directly.
Frequently asked questions
How does Amazon Bedrock AgentCore pricing work?
AgentCore bills each service independently. Runtime charges $0.0895 per vCPU-hour and $0.00945 per GB-hour based on active resource use. Because agents spend 30 to 70 percent of session time waiting on LLM responses or tool calls, idle CPU accrues no charge. You pay only for active consumption. Gateway charges per tool invocation. Memory charges per event stored and per retrieval. Each component can be used independently or together.
Does the AgentCore cost estimate include LLM model inference?
This calculator does. Many AgentCore cost tools estimate infrastructure only and ask you to add Bedrock inference separately. This calculator includes model inference as a line item, using the current AWS pricing for the model you select. For many deployments, model costs represent 50 to 70 percent of total spend, but this depends on the model you're using.
What's the difference between AgentCore pricing and Bedrock pricing?
Bedrock pricing covers foundation model inference, or what you pay per token to run Claude, Nova, Llama, or another model. AgentCore pricing covers the agent infrastructure around that model: the runtime environment, session isolation, memory, tool gateway, and observability. You incur both on any agent running in production. This calculator estimates both together.
What isn't included in my estimate?
Four things fall outside this estimate: Bedrock Prompt Caching savings (can reduce input token costs significantly for agents with repeated system prompts), data transfer charges at EC2 standard rates, ECR storage for container-deployed agents, and the cost of custom model overrides in Memory extraction strategies. For most POC and early-production workloads these are minor. At scale, Prompt Caching in particular is worth modeling separately.
When does the cost get complicated enough to need help?
For simple agents with predictable session volume, the estimate here is close enough to build a business case. Cost gets harder to model when you add persistent per-user memory, multi-agent orchestration, high tool call frequency, or RAG retrieval. Those architectures have compounding cost paths that depend heavily on design choices. If you're sizing a production deployment and the number matters to stakeholders, Tech 42's AgentCore Accelerate Program includes architecture review and AWS cost modeling as part of the two-week engagement, often funded through AWS.
