Amazon Bedrock AgentCore

AgentCore is the layer AWS added around Bedrock for running agents in production: serverless execution, memory, tool routing, identity and traces. It went generally available on 13 October 2025. The thing to understand before you design around it is that it is billed entirely separately from model tokens — these are two independent meters on the same workload, and AWS prices them on two different pages.

This page adds them up, and shows which components you can actually get in your region.

What an agent actually costs

A worked example on Claude Sonnet 4.6: 500 agent runs a day, 6 model steps each, 12K input and 800 output tokens per step, 70% prompt-cache hit rate, 4s of active CPU per run.

Line Monthly
Model input tokens $1,199
Model output tokens $1,080
Runtime — CPU $1.49
Runtime — memory $0.32
Gateway invocations $0.30
Memory events $7.50
Total per month $2,288

The same workload on three models

Identical agent, identical traffic — only the model changes. The AgentCore column is almost flat; the total is not.

Model AgentCore Total / month AgentCore share
Nova Micro $9.61 $37.64 25.5%
Claude Sonnet 4.6 $9.61 $2,288 0.4%
Claude Fable 5 $9.61 $7,606 0.1%

Which is the practical conclusion: if the agent bill is too high, the model is almost always the lever, not the platform. Prompt caching is the second — see models that support it.

Where the platform bill does bite

  • Web Search at $7 per 1,000 queries. One search per step on this workload adds $630.00 a month — more than the model. Search, not tokens, becomes the largest line.
  • Policy input tokens at $0.13 per 1,000. That is $130 per 1M tokens, well above most model rates on this site. Keep policy prompts short.
  • A browser session left open. Built-in Tools bill on the same vCPU-hour meter as Runtime, but a browser does not stop consuming when the agent stops thinking.
  • No prompt caching. This example assumes a 70% cache hit rate. At 0% the token line roughly triples, and nothing about AgentCore changes that.

AgentCore by region

Runtime, Gateway and Identity reach 15 of the 31 regions this site tracks. The exception is US West (N. California), Canada West (Calgary), Mexico (Central), Europe (Zurich), EU (Milan), Europe (Spain), Asia Pacific (Taipei), Asia Pacific (Osaka), Asia Pacific (Hyderabad), Asia Pacific (Jakarta), Asia Pacific (Melbourne), Asia Pacific (Malaysia), Asia Pacific (New Zealand), Asia Pacific (Thailand), Israel (Tel Aviv), Africa (Cape Town) — if your model runs there, your agent runtime cannot, and you are making a cross-region call whether you planned to or not.

Region Runtime Gateway Memory Identity Built-in Tools AgentCore Evaluations Observability
us-east-1 US East (N. Virginia) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
us-east-2 US East (Ohio) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
us-west-1 US West (N. California) Not available Not available Not available Not available Not available Not available Not available Not available
us-west-2 US West (Oregon) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
ca-central-1 Canada (Central) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
ca-west-1 Canada West (Calgary) Not available Not available Not available Not available Not available Not available Not available Not available
mx-central-1 Mexico (Central) Not available Not available Not available Not available Not available Not available Not available Not available
eu-west-1 EU (Ireland) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
eu-west-2 EU (London) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
eu-west-3 EU (Paris) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
eu-central-1 EU (Frankfurt) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
eu-central-2 Europe (Zurich) Not available Not available Not available Not available Not available Not available Not available Not available
eu-north-1 EU (Stockholm) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
eu-south-1 EU (Milan) Not available Not available Not available Not available Not available Not available Not available Not available
eu-south-2 Europe (Spain) Not available Not available Not available Not available Not available Not available Not available Not available
ap-east-2 Asia Pacific (Taipei) Not available Not available Not available Not available Not available Not available Not available Not available
ap-northeast-1 Asia Pacific (Tokyo) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
ap-northeast-2 Asia Pacific (Seoul) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
ap-northeast-3 Asia Pacific (Osaka) Not available Not available Not available Not available Not available Not available Not available Not available
ap-south-1 Asia Pacific (Mumbai) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
ap-south-2 Asia Pacific (Hyderabad) Not available Not available Not available Not available Not available Not available Not available Not available
ap-southeast-1 Asia Pacific (Singapore) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
ap-southeast-2 Asia Pacific (Sydney) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand
ap-southeast-3 Asia Pacific (Jakarta) Not available Not available Not available Not available Not available Not available Not available Not available
ap-southeast-4 Asia Pacific (Melbourne) Not available Not available Not available Not available Not available Not available Not available Not available
ap-southeast-5 Asia Pacific (Malaysia) Not available Not available Not available Not available Not available Not available Not available Not available
ap-southeast-6 Asia Pacific (New Zealand) Not available Not available Not available Not available Not available Not available Not available Not available
ap-southeast-7 Asia Pacific (Thailand) Not available Not available Not available Not available Not available Not available Not available Not available
il-central-1 Israel (Tel Aviv) Not available Not available Not available Not available Not available Not available Not available Not available
af-south-1 Africa (Cape Town) Not available Not available Not available Not available Not available Not available Not available Not available
sa-east-1 South America (Sao Paulo) On-demand On-demand On-demand On-demand On-demand On-demand On-demand On-demand

Transcribed from the AWS region support table, verified 2026-08-04. AgentCore has no public API for this, so unlike model availability it is not machine-derived — see methodology.

The components, and what each bills

You can adopt these independently — Runtime without Memory, Gateway without Runtime. Prices verified 2026-08-04 against the AWS pricing page.

AgentCore Runtime

Serverless execution with per-session isolation

Runs your agent as a managed, serverless workload. Each session gets its own isolated microVM, and you are billed for CPU only while the agent is actually computing — not while it waits on a model response or an HTTP call.

Agents spend most of their wall-clock time waiting. Billing CPU on active use rather than session duration is the difference between paying for thinking and paying for waiting, and it is the single biggest reason Runtime is cheaper than holding a container open.

per vCPU-hour
$0.089
per GB-hour
$0.009

AgentCore Gateway

Turns APIs and Lambdas into MCP tools

Exposes existing REST APIs, Lambda functions and third-party services to your agent as Model Context Protocol tools, with a semantic search index so the agent finds the right tool instead of being handed all of them.

Tool descriptions live in the prompt, so every unused tool is tokens you pay for on every single step. Semantic tool selection keeps that prefix small, which compounds hard in a loop.

per 1,000 API invocations
$0.005
per 1,000 search queries
$0.025
per 100 tools indexed / month
$0.020

AgentCore Memory

Short-term and long-term recall

Managed conversation state. Short-term memory holds the working session; long-term memory extracts durable facts and retrieves them into later sessions.

Priced per event and per record, not per token — so unlike stuffing history into the prompt, cost does not scale with how long the conversation gets.

per 1,000 new short-term events
$0.250
per 1,000 long-term records stored / month
$0.750
per 1,000 long-term retrievals
$0.500

AgentCore Identity

OAuth and credentials on the agent's behalf

Brokers tokens so an agent can act against Google, Slack, GitHub or your own OAuth provider as a specific user, without the credentials passing through your prompt or your code.

Free when the agent reaches tools through Runtime or Gateway. Route authentication that way and this line disappears from the bill entirely.

per 1,000 token or API-key requests
$0.010

No charge when used through AgentCore Runtime or Gateway.

Built-in Tools

Browser and Code Interpreter

A managed headless browser for agents that operate real web UIs, and a sandboxed code interpreter for agents that need to compute rather than estimate.

Both bill on the same vCPU/GB-hour meter as Runtime, so a browser session left open costs the same as an agent left thinking. Close them.

per vCPU-hour
$0.089
per GB-hour
$0.009

Policy in AgentCore

Rules on what the agent may do

Natural-language and structured authorization rules evaluated before an agent takes an action, so 'never issue a refund over €500 without approval' is enforced outside the model rather than asked of it.

The token component is billed separately from your model spend, and at $0.13 per 1,000 input tokens it is well above most model rates — keep policy prompts short.

per authorization request
$0.000025
per 1,000 input tokens
$0.130

AgentCore Evaluations

Scoring agent trajectories

Judges whether an agent actually did the right thing across a whole multi-step trajectory, rather than whether one response looked plausible.

The failure mode of agents is a sequence that goes wrong in step four, which single-response evaluation cannot see.

per 1,000 built-in input tokens
$0.002
per 1,000 built-in output tokens
$0.012
per 1,000 custom evaluations
$1.50

AgentCore Observability

Traces of every step

OpenTelemetry-compatible traces of the agent's reasoning, tool calls and token usage, surfaced in CloudWatch.

Billed at CloudWatch rates rather than a separate AgentCore meter — the cost lands on a different line of your bill than the rest of this table.

Charged at standard CloudWatch rates; there is no separate AgentCore price.

Web Search

Managed search for agents

A first-party search tool an agent can call without you contracting a third-party search API.

At $7 per 1,000 queries this is the most expensive per-call item in AgentCore by a wide margin. An agent that searches on every step will make search, not tokens, the largest line on the bill.

per 1,000 queries
$7.00

Which model to run on it

AgentCore is model-agnostic, but an agent loop punishes two specific weaknesses: no tool use means it cannot drive the loop at all, and no prompt caching means every step re-pays full price for the system prompt. These are the cheapest models on this site that do tool use, with their caching support called out.

Model Input /1M Caching
Nova Micro Amazon $0.035 yes
Nova Lite Amazon $0.060 yes
GPT-5.6 Luna OpenAI $0.200 yes
Llama 3.1 8B Instruct Meta $0.220 no
Claude 3 Haiku Anthropic $0.250 no
Llama 3.1 70B Instruct Meta $0.720 no
Llama 3.3 70B Instruct Meta $0.720 no
Nova Pro Amazon $0.800 yes

Model the whole thing on your own numbers in the cost calculator, or let the picker shortlist one from six questions.

Questions

Is AgentCore included in Bedrock model pricing?

No. AgentCore is billed separately and on top of model tokens. Running an agent on AgentCore Runtime with Claude Sonnet 5 means paying the Sonnet token rate for every model call and the Runtime vCPU/GB-hour rate for the compute around it. The two meters are independent.

Does AgentCore only work with Bedrock models?

No. AgentCore Runtime is model-agnostic and framework-agnostic — it will run an agent built on LangGraph, CrewAI or Strands Agents, calling a model hosted outside Bedrock. Using Bedrock models keeps everything on one bill and inside one region, which is the practical reason most teams do.

What does AgentCore actually cost for a real agent?

For most agents the token bill still dominates. Runtime bills CPU only while the agent computes, so a step that spends four seconds waiting on the model and 200ms deciding is charged for the 200ms. The items that surprise people are Web Search at $7 per 1,000 queries and Policy input tokens at $0.13 per 1,000 — both far above typical model rates.

Can I use AgentCore in the same region as my model?

Usually, but not always, and the mismatch is the thing to check before you design around it. AgentCore Runtime, Gateway and Identity cover every region this site tracks except Asia Pacific (Osaka). If your model runs somewhere AgentCore does not, you are making a cross-region call, with the latency and data-residency consequences that implies.

Do I need AgentCore to build an agent on Bedrock?

No. You can call the Bedrock Converse API with tool use from your own service and own the loop yourself. AgentCore is worth paying for when session isolation, managed memory, credential brokering or per-step traces would otherwise be code you write and operate.