AgentCore pricing cheat sheet

Every meter, plus two worked sketches at startup scale. Built from Lesson 2. Rates from the AgentCore pricing page. Verified 2026-08-05. Consumption-based throughout, with no upfront commitment and no minimum fees.

Compute rarely makes an AgentCore bill surprising. Per-unit meters do — Memory retrievals and Web Search especially. Find the meter their agent hits most often per turn, not the one with the biggest rate.

All meters

Rates as published 2026-08-05
CapabilityMeterRate
Runtime, Browser, Code Interpreter vCPU-hour$0.0895
GB-hour$0.00945
Gateway API invocations (ListTools, InvokeTool, Ping)$0.005 / 1k
Search API$0.025 / 1k
Tool indexing$0.02 / 100 tools / mo
Identity Token or API key requests, non-AWS resources$0.010 / 1k
Used via Runtime or GatewayNo charge
Memory Short-term, new events$0.25 / 1k
Long-term storage, built-in strategy$0.75 / 1k records / mo
Long-term storage, built-in w/ overrides or self-managed$0.25 / 1k records / mo
Long-term retrieval$0.50 / 1k
Policy Authorization request$0.000025 each
Input tokens processed$0.13 / 1k
Evaluations Built-in evaluators, input / output tokens$0.0024 / $0.012 per 1k
Custom evaluators$1.50 / 1k (model separate)
Batch evaluations, input / output$0.0018 / $0.009 per 1k
Web Search Queries$7.00 / 1k
Agent Registry RecordsFirst 5k/mo free, then $0.400 / 1k
Search APIFirst 1M/mo free, then $0.020 / 1k
List and GetFirst 2M/mo free, then $0.004 / 1k
Payments CreateInstrument / ProcessPayment Wallet provider pass-through
harness, AgentCore CLI, skills n/aNo charge
Optimization Preview Insights / recommendations / A B tests Free in preview; pay only underlying resources
Cross-region inference n/aNo charge

Memory has three meters, not one

This is the single most useful thing on this card. Memory bills separately for events written, records stored, and records retrieved. Teams model the first two and forget the third, but a chatty agent retrieves far more often than it writes, so retrieval routinely becomes the largest line in the whole AgentCore bill.

The strategy choice also moves cost on stored records: $0.75 per 1k/month on a built-in strategy versus $0.25 on built-in with overrides or self-managed. That cheaper option is the same one that avoids cross-region inference, so for a Canadian customer it is a rare case where the compliance-friendly choice is also the cheap one. Lead with that.

Worked sketch A: customer support agent

Illustrative only. Assumptions stated so a customer can argue with them. For a real number, use the AWS Pricing Calculator.

AssumptionValue
Sessions per month10,000
Turns per session8
Active compute per session90 seconds
Runtime sizing1 vCPU / 2 GB
Tool calls per turn1.5
Memory retrievals per turn2
Long-term records held15,000
LineWorkingMonthly
Runtime vCPU250 vCPU-hr × $0.0895$22.38
Runtime memory500 GB-hr × $0.00945$4.73
Gateway invocations130k × $0.005/1k$0.65
Gateway tool indexing40 tools$0.02
Identityvia Runtime and Gateway$0.00
Memory, events80k × $0.25/1k$20.00
Memory, storage (built-in)15k × $0.75/1k$11.25
Memory, retrieval160k × $0.50/1k$80.00
Total≈ $139

Read the shape, not the total

Memory is 80% of this bill and retrieval alone is 58% of it, nearly three times the entire Runtime cost. Switching to a built-in-with-overrides strategy saves $7.50 on storage; halving retrievals per turn saves $40. Retrieval discipline is the optimisation that matters, and it is an application design decision, not a configuration one.

Worked sketch B: web research agent

AssumptionValue
Research tasks per month2,000
Browser time per task12 minutes @ 2 vCPU / 4 GB
Code Interpreter per task3 minutes @ 1 vCPU / 2 GB
Runtime orchestration per task15 minutes @ 1 vCPU / 2 GB
Web Search queries per task15
LineWorkingMonthly
Browser800 vCPU-hr + 1,600 GB-hr$86.72
Code Interpreter100 vCPU-hr + 200 GB-hr$10.84
Runtime500 vCPU-hr + 1,000 GB-hr$54.20
Web Search30k queries × $7.00/1k$210.00
Total≈ $362

Different agent, different dominant meter

Web Search is 58% of this bill and all the compute together is 42%. At $7.00 per 1,000 queries it is the most expensive per-unit meter in AgentCore by a wide margin. For any research or browsing agent, query discipline (caching, deduplication, narrowing before searching) is the entire cost conversation.

How to run this conversation with a founder

  1. Ask what the agent does per turn, not how many users they have. Turns and per-turn tool and retrieval counts drive everything.
  2. Find the dominant meter first. Chatty stateful agent → Memory retrieval. Research or browsing agent → Web Search. Long-running autonomous agent → vCPU-hours. Everything else is rounding.
  3. Name the free things. harness, the CLI, skills, Identity via Runtime or Gateway, cross-region inference, Optimization in preview, and Registry's free tiers. This is a genuinely generous list and it reframes the conversation.
  4. Send them to the calculator. Do not leave a number in their notes that came from you. Send them to calculator.aws.

Two honest caveats on these numbers

The Policy token meter needs care. Policy publishes $0.000025 per authorization request and $0.13 per 1,000 input tokens processed, and the docs do not say which configurations incur which. A content-heavy setup could dominate a bill by an order of magnitude. Verify with the service team before modelling Policy at scale.

No independent benchmark exists comparing AgentCore against self-managed ECS or EKS for the same workload, so any claim either way is unproven. Argue Runtime on session isolation and undifferentiated engineering removed, not on unit cost.