Amazon Bedrock AgentCore
The broadest production-agent platform on the table and the only one that speaks both MCP and A2A across the stack — but there is no platform fee to anchor a budget on, thirteen separate meters underneath, and it runs in nine AWS regions or nowhere.
PRICING
| Harness | $0 | The managed agent loop carries no separate charge; you pay for the resources it consumes |
| Runtime microVMs | $0.0895/vCPU-hr + $0.00945/GB-hr | Per-second billing, 1-second minimum, 128MB memory floor. CPU billed on active cycles only; memory billed on peak for the full session. AWS's own worked example: a 60-second support session with 70% I/O wait costs $0.0007235, so 10M sessions per month is $7,235 |
| Runtime instances | EC2 price + management fee | For persistent or resource-heavy agents. EC2 runs in your account at your rates and your Savings Plans apply to the compute but not the management fee. Sessions persist up to 14 days |
| Browser / Code Interpreter | $0.0895/vCPU-hr + $0.00945/GB-hr | Same active-consumption rates. AWS's browser example: 100K sessions of 10 minutes at 80% I/O wait is $1,226.67 per month |
| Web Search | $7.00 per 1,000 queries | GA 17/06/2026. Runs as a Gateway MCP connector with zero data egress from your AWS environment |
| Gateway, Memory, Identity, Policy, Evaluations, Observability | Separate meters | Request, storage, token and trace meters, each enabled independently. Observability log ingestion is the line most often underestimated |
| Foundation model tokens | Amazon Bedrock rates | Billed separately per token. AgentCore does not replace or discount Bedrock inference pricing |
There is no subscription to compare and no seat price — thirteen independent meters and a model bill on top. The published $0.0895 and $0.00945 rates were verified on the AWS pricing page on 04/09/2026; that page changes more often than most write-ups about it. The advertised saving is real for CPU, because agent sessions sit in I/O wait 30-70% of the time and idle CPU is not billed, but memory bills on peak for the entire session including that wait — a long idle session is not a free one.
checked 2026-09-04 · vendor pricing page
Element scores
Strengths
Nothing else on this table covers as much of the agent stack at this quality. Since GA on 13/10/2025 AWS has shipped continuously and the additions are the ones production teams actually block on: Policy went GA on 03/03/2026 and turns plain-language rules into Cedar enforced at Gateway in milliseconds; Evaluations went GA on 31/03/2026 with 13 built-in evaluators running both as continuous sampling of live traffic and as on-demand regression tests in CI; Web Search and Bedrock Managed Knowledge Base both landed on 17/06/2026, closing the current-facts and own-documents gaps in one day; AWS Agent Registry went GA in August 2026 with organisation-wide auto-discovery and cross-account sharing. Protocol support is the strongest we have scored — Gateway consumes and publishes MCP, Runtime speaks A2A, and AWS has demonstrated a Google ADK orchestrator driving Strands and OpenAI SDK agents together on one runtime. The framework-agnostic claim holds up: CrewAI, LangGraph, LlamaIndex, Google ADK, OpenAI Agents SDK and Strands all deploy, and models can come from outside Bedrock entirely. Observability is OTEL-compatible and exports to Datadog, Dynatrace, LangSmith and Langfuse instead of trapping you in CloudWatch. The service picked up ISO and CSA STAR certification in February 2026, and the platform got measurably faster that month too: sequential calls 25-35% quicker, evaluation P90 down 37-50%.
Honest dings
You cannot budget this from the pricing page. There are thirteen meters, no platform fee to anchor on, and the foundation-model tokens — usually the largest line — are billed somewhere else entirely. AWS's own worked examples are honest but small: $0.0007 per session reads like nothing until you notice their own illustration reaches $7,235 a month at 10M sessions, and that figure excludes Gateway, Memory, Policy, Evaluations, Observability and every token. Independent cost breakdowns consistently name the same trap — teams model Runtime and forget Memory extraction, Gateway tool volume and observability log ingestion, with one worked example putting a 50K-session month at roughly $511 of AgentCore charges of which $358 was observability logs alone. The 'you don't pay for I/O wait' pitch is true for CPU and not for memory: peak memory bills for the full session lifetime, so an agent parked waiting on a slow tool still accrues. Geography is a hard limit — nine regions, with Frankfurt and Ireland the only European options, so anything needing data resident elsewhere in the EU is out. And while models and frameworks are genuinely portable, the operational layer is not: Gateway targets, Policy in Cedar, Memory strategies, Identity wiring and Registry records are AWS constructs that do not travel. Free-first buyers should note the honest comparison is not AgentCore versus nothing but AgentCore versus a self-hosted stack that costs the same as the box it runs on.
Sources (14) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- aws.amazon.com/bedrock/agentcore/pricing — Live official pricing verified 04/09/2026: Runtime microVM / Browser / Code Interpreter at $0.0895 per vCPU-hour and $0.00945 per GB-hour, Web Search $7.00 per 1,000 queries, Harness free of extra charge, $200 Free Tier credits; AWS's own worked examples — 10M support sessions at $7,235/month and 100K browser sessions at $1,226.67/month (accessed 2026-09-04)
- docs.aws.amazon.com/bedrock-agentcore/latest/devgui… — Official service catalogue: Harness, Runtime, Memory, Gateway, Identity, Code Interpreter, Browser, Observability; framework support CrewAI, LangGraph, LlamaIndex, Google ADK, OpenAI Agents SDK, Strands; MCP and A2A protocol support; models in and outside Bedrock incl. OpenAI and Gemini (accessed 2026-09-04)
- aws.amazon.com/about-aws/whats-new/2025/10/amazon-b… — Official GA announcement 13/10/2025: nine regions listed, VPC / PrivateLink / CloudFormation / tagging across all services, eight-hour Runtime windows with session isolation, A2A on Runtime, Gateway connecting to existing MCP servers with IAM plus OAuth, OTEL-compatible Observability exporting to Dynatrace, Datadog, Arize Phoenix, LangSmith and Langfuse (accessed 2026-09-04)
- aws.amazon.com/blogs/aws/amazon-bedrock-agentcore-a… — Official: Policy GA 03/03/2026 and Evaluations GA 31/03/2026; AgentCore SDK downloaded over 2 million times in five months since preview; built-in evaluator dimensions and continuous live sampling (accessed 2026-09-04)
- aws.amazon.com/about-aws/whats-new/2026/03/agentcor… — Official Evaluations detail: 13 built-in evaluators, online continuous plus on-demand CI/CD regression modes, Ground Truth reference answers and behavioural assertions, custom LLM-based and Lambda code-based evaluators, nine regions (accessed 2026-09-04)
- docs.aws.amazon.com/bedrock-agentcore/latest/devgui… — Official release notes: ISO and CSA STAR certification and 25-35% Runtime latency cut plus 37-50% evaluation P90 improvement (Feb 2026); Agent Registry GA with auto-detection, KMS CMK, PrivateLink and RAM cross-account sharing (Aug 2026); Memory IngestData, flexible namespaces and 100KB JSON payloads (Aug 2026); Identity Consent Portal and TypeScript evaluation support (Sep 2026) (accessed 2026-09-04)
- aws.amazon.com/blogs/aws/announcing-web-search-on-a… — Official Web Search GA: built-in Gateway connector target over MCP, returns snippets with source URLs, titles and publication dates, zero data egress from the customer's AWS environment (accessed 2026-09-04)
- aws.amazon.com/blogs/machine-learning/introducing-a… — Official A2A walkthrough: peer discovery and capability sharing across frameworks, demonstrated with a Google ADK orchestrator coordinating Strands and OpenAI SDK agents on AgentCore Runtime with Cognito OAuth 2.0 machine-to-machine auth (accessed 2026-09-04)
- cloudburn.io/blog/amazon-bedrock-agentcore-pricing — Independent 12-component cost breakdown reproducing the official CPU and memory formulas, and a worked 50K-session month totalling roughly $511 of AgentCore charges of which observability log ingestion is $358 (accessed 2026-09-04)
- cipherprojects.com/blog/posts/amazon-bedrock-agentc… — Independent consultancy view: the recurring budgeting failure is modelling Runtime while forgetting Memory extraction, Gateway tool volume and per-turn model tokens; memory bills on peak footprint for the session lifetime even while idle (accessed 2026-09-04)
- hidekazu-konishi.com/entry/amazon_bedrock_agentcore… — Independent production operations guide (verified 26/04/2026): confirms the same unit prices across Runtime, Code Interpreter and Browser, warns the pricing page changes faster than write-ups, and notes idle I/O wait still bills at the in-flight resource shape (accessed 2026-09-04)
- caylent.com/blog/what-amazon-bedrock-agentcores-new… — Independent partner analysis of the June 2026 wave — managed harness, Managed Knowledge Base, Web Search — and the cost-composition shift: less custom development traded for new usage-based meters, integration work and evaluation workloads (accessed 2026-09-04)
- rackspace.com/blog/amazon-bedrock-agentcore-preview… — Independent hands-on review: Gateway adds tools without restarts or code changes, Memory uses namespaces with semantic filtering and top-K retrieval; flags limited visibility into memory contents as a gap (accessed 2026-09-04)
- awsinsider.net/articles/2025/10/23/amazon-bedrock-a… — Independent GA coverage (23/10/2025) confirming the composable-service model, eight-hour execution windows, A2A on Runtime, Gateway MCP transformation with IAM authorization, and nine-region availability (accessed 2026-09-04)