by Amazon Web Services
Amazon Bedrock Managed Agents, powered by OpenAI review, pricing and verdict
OpenAI models and the Codex agent loop running stateful agents inside AWS.
- Best for
- Cloud architects at AWS-first enterprises
- Not for
- Solo builders who want a flat monthly plan
Should you use Amazon Bedrock Managed Agents, powered by OpenAI?
Pick it if your agents must live inside an AWS account with per-agent IAM identities and CloudTrail logs, and you want OpenAI's Codex agent loop driving the work. It is a preview with text-only input and no subagents, so treat it as a pilot. The AgentCore example caps idle time and lifetime at 8 hours.
Runs OpenAI's agent loop with a per-session AWS IAM role, while model inference and the agent runtime stay inside AWS.
The AWS channel, existing commitments and AgentCore integration give it a distribution edge, but the same OpenAI models and agent loop are sold by OpenAI directly, and the feature set is still a preview.
Also for
- Platform engineers running agents on AgentCore
- Cloud architects at AWS-first enterprises that want OpenAI agents under existing IAM and CloudTrail controls
Show 3 more
- Platform engineers building internal coding and document-processing agents on AgentCore
- Security leads who need per-agent identity and an audit trail before approving agents
- Teams already using OpenAI Codex that want billing and governance consolidated on AWS
Also not for
- Teams that need EU data residency or multimodal input today
Main trade-off
It is a preview limited to three US Regions with text input and no subagents.
Amazon Bedrock Managed Agents, powered by OpenAI pricing
Server plans
| Plan |
|---|
| Self-hosted computeExecution environment · Your own host, workspace and network with a running exec server. Suits an existing dev machine or container. · US East (N. Virginia) · US West (Oregon) · US East (Ohio) |
| Amazon Bedrock AgentCore RuntimeExecution environment · A Runtime holding the exec server plus storage and networking in your AWS account. The example sets idle and maximum lifetime at 28,800 seconds. · US East (N. Virginia) · US West (Oregon) · US East (Ohio) |
- Minimum commitment
- None published for the preview
- Billing
- Pay as you go · Committed spend
How the pricing works, in full
AWS lists no separate fee for managed agents in the preview docs: you pay for model inference on Amazon Bedrock and for the AWS resources your application uses. Amazon states that OpenAI model pricing on Bedrock matches OpenAI first-party rates, counts toward existing AWS commitments and carries a 90% cached-input discount. The AgentCore example creates storage and a NAT gateway that keep billing when no turn is running. Preview pricing can change.
Costs that are easy to miss
- The AgentCore example deploys storage and a NAT gateway that AWS says can keep incurring charges when no turn is running.
- Your own execution environment, whether a host or a Runtime, bills separately from model inference.
Amazon Bedrock Managed Agents, powered by OpenAI regions and compliance
Available regions
- US East (N. Virginia)
- US West (Oregon)
- US East (Ohio)
Managed services
- Amazon Bedrock AgentCore Runtime
- Codex exec server
- Filesystem skills
- STDIO MCP servers
- IAM session roles
- Session event streaming
Access methods
- API
Networking
- Private networksSupported
What Amazon Bedrock Managed Agents, powered by OpenAI does
Amazon Bedrock Managed Agents, powered by OpenAI, is an AWS public preview that runs stateful agents on OpenAI models through bedrock-mantle endpoints in 3 US Regions. A session holds the model, instructions, tools and IAM role, while commands execute on compute you supply: your own host or Amazon Bedrock AgentCore Runtime.
Stateful sessions
A session holds the model, instructions, tools, IAM role and execution environment, and retains conversation context across messages.
Codex agent loop at the core
AWS says the runtime pairs OpenAI models with the Codex agent loop to keep context across long, multi-step assignments.
Bring your own compute
Commands and local tools run on self-hosted compute or an AgentCore Runtime that connects to the service through an outbound exec-server link.
Skills and MCP tools
The agent discovers filesystem-based skills and uses environment-based STDIO MCP servers, with a cap of 32 unique capability-directory paths.
Show all 6 features
Per-agent AWS identity
Each session assumes a same-account IAM role passed with iam:PassRole, and requests are signed with SigV4 against the bedrock-mantle service.
Durable items and event stream
Conversation output is stored as items you can page through (1 to 100 per page, default 20), while an event stream reports progress as turns run.
Strengths and limits
Strengths
- Keeps the agent runtime and model inference inside AWS, so identity, permissions and logging stay on controls a security team already operates.
- Execution stays on compute you choose, which means files, credentials and network reach are yours to restrict rather than a vendor's sandbox.
- Needs no OpenAI API key: the preview examples authenticate with standard AWS credentials.
Limits
- Public preview with APIs that can change, text-only input and no subagents, so it is a pilot target rather than a production default today.
- Only three US Regions are supported, and model availability differs by Region and account, which rules out EU or APAC data residency for now.
- AWS publishes no separate pricing page for it and no compliance attestations for the preview on the pages reviewed, so procurement has to ask AWS directly.
Amazon Bedrock Managed Agents, powered by OpenAI alternatives
| Service | Regions | Pricing | Compliance | Checked |
|---|---|---|---|---|
| Amazon Bedrock Managed Agents, powered by OpenAI This page | US | Pay as you go · Committed spend | -- | 30 Sep |
| Claude in Amazon BedrockPick Claude in Amazon Bedrock if you want Anthropic models under the same AWS controls, with 27 Regions listed on its record. | Africa · Asia Pacific · Canada · Europe +3 | Pay as you go · Committed spend | HIPAA · GDPR · FedRAMP | 10 Oct |
 Google Vertex AIPick Google Vertex AI if your data and identity already sit on Google Cloud rather than AWS. | US · Canada · South America · Africa +3 | Free tier · Pay as you go · Committed spend | ISO 27001 · SOC 2 · HIPAA | 26 Oct |
 BasetenPick Baseten if you want to serve your own or open models instead of a managed OpenAI agent runtime. | US · Europe · Asia Pacific | Free tier · Pay as you go · Hourly | SOC 2 · HIPAA · GDPR | 29 Oct |
Amazon Bedrock Managed Agents, powered by OpenAI facts
- Provider founded
- 2006
- Headquarters
- Seattle, WA, USA
- Open source
- No
What changed
- Release Amazon Bedrock Managed Agents, powered by OpenAI, announced in limited preview source
Questions about Amazon Bedrock Managed Agents, powered by OpenAI
What are the pricing plans for Amazon Bedrock Managed Agents in 2026?
AWS's preview documentation lists no separate per-agent fee. You pay for model inference on Amazon Bedrock plus the AWS resources your application runs, and Amazon says OpenAI models on Bedrock are priced at OpenAI first-party rates with a 90% cached-input discount. The AgentCore example adds storage and a NAT gateway that can bill while idle, so delete the stack after testing.
Is Bedrock Managed Agents available in every AWS Region?
No. The preview runs on bedrock-mantle endpoints in US East (N. Virginia), US West (Oregon) and US East (Ohio) only. Cross-Region inference profiles are not supported in this workflow, and the models you can use depend on your Region and account.
What are the best alternatives to Amazon Bedrock Managed Agents?
For Anthropic models under AWS controls, use Claude in Amazon Bedrock. For a Google Cloud stack, Google Vertex AI covers the same buyer. If you would rather call OpenAI directly, the OpenAI Agents API skips the AWS wrapper but also the IAM and CloudTrail integration.
Bedrock Managed Agents or AgentCore: which should you pick?
They stack rather than compete. AgentCore is the open platform for agents on any model and framework, and it supplies the default compute for managed agents. Choose managed agents when you want OpenAI's agent loop handling the session, and plain AgentCore when you bring your own model or framework.
What does it take to start using Bedrock Managed Agents?
You need an AWS account in a supported Region, a session role that trusts bedrock-mantle.amazonaws.com, and iam:PassRole on that role for the caller. Requests are signed with SigV4 and no OpenAI API key is required. AWS supplies an example bundle with a self-hosted path and an AgentCore CDK path for the first run.
About Amazon Bedrock Managed Agents, powered by OpenAI
Amazon Bedrock Managed Agents, powered by OpenAI, is a preview service from Amazon Web Services and OpenAI that runs stateful agents on OpenAI models without leaving an AWS account. AWS describes it as built jointly with OpenAI, combining the Codex agent loop with Amazon Bedrock AgentCore. You create a session, give it a model, instructions, tools and an IAM role, then submit messages. The service keeps the conversation and the model calls, and your code runs on compute you provide.
Read the full background
That split is the main design choice. Tools and shell commands execute in an execution environment you own: either your own host or an AgentCore Runtime that holds the Codex exec server. The exec server makes an outbound connection to the service, so the agent's files and the conversation have separate lifecycles. Deleting a session does not delete anything in your S3 buckets or on your host. Skills load from the filesystem and additional tools arrive through STDIO MCP servers running in that environment.
The service sits next to other OpenAI offerings on AWS. OpenAI Codex and the GPT line are sold on Bedrock as models, and AWS's own example policy names GPT-5.6 Luna. The catalog now includes GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna on Bedrock in general. The managed agents docs only promise OpenAI models that are available in your Region and account, so check model discovery before you design around one.
It is a public preview and the limits are real. Input is text, subagents are not supported, there is no built-in long-term memory, cross-Region inference profiles do not work, and the service-managed session data has no customer-managed KMS key. If you want Anthropic's models inside the same AWS boundary, look at Claude in Amazon Bedrock. If you want a Google-side equivalent, compare Google Vertex AI. OpenAI's own hosted route is the OpenAI Agents API.
The buyer is an AWS-first team that already runs identity, networking and audit on IAM, PrivateLink and CloudTrail and wants agents under those controls. Teams that need a finished product with long-term memory, multimodal input or a console workflow should wait for general availability, because the preview procedures use APIs and supplied deployment examples rather than a console.
Sources and guides
Official sources
Plans, regions, the uptime commitment and the compliance list on this page are read from Amazon Web Services's own pages and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Hokai guides covering Amazon Bedrock Managed Agents, powered by OpenAI
Lists that include Amazon Bedrock Managed Agents, powered by OpenAI
More AI Infrastructure Services on Hokai
LuminalCompiler-built inference for open models, as a serverless API or on-prem.