OpenAI Codex on Amazon Bedrock: AWS Maps Three Rollout Paths
AWS published an enterprise guide for running Codex inference through Bedrock, with a least-privilege pilot and an open source quick-start repository.
What AWS published
On 11 November 2026 the AWS Machine Learning Blog published a practical guide to OpenAI Codex on Amazon Bedrock, written by Shivi Bhatia, Sebastien Couturier and Sriharsha M S. The setup described lets Codex use Bedrock for inference with supported OpenAI models, while Codex itself keeps running on the developer's workstation. The problem the post addresses is organisational as much as technical: platform teams want developers using these tools without building a second stack for model access, identity and operations. The guide is about fitting the tool into the IAM, Regional and observability practices an organisation already runs.
The worked examples use GPT-5.6 Sol. Before any pilot, AWS asks readers to check its supported-models-by-Region page and confirm that Sol, Terra or Luna is available in the Region and inference path they plan to use.
Three ways to route the requests
The post frames the main design choice as how model requests reach Bedrock, and where policy and operational controls get applied. It offers three paths:
- Native AWS Access. The recommended default for a small, least-privilege pilot when AWS identity and IAM controls are enough. Employees sign in through the workforce identity provider they already use; in the example, Okta handles authentication over SAML and provisions users and groups into IAM Identity Center over SCIM, which then issues temporary AWS credentials to Codex. The identity provider is not configured inside Codex.
- AgentCore Gateway. A capability of Amazon Bedrock AgentCore, aimed at teams that want AWS-managed gateway infrastructure for routing, Amazon Bedrock Guardrails or related controls.
- An LLM Gateway. For organisations that already operate a gateway or want controls enforced in the request path. AWS describes a custom integration as the advanced version of this pattern rather than a quick start.
Supporting files come from an open source repository, Guidance for Codex on AWS, which ships quick starts for Native AWS Access, AgentCore Gateway, LiteLLM on ECS, Portkey, the AWS Distro for OpenTelemetry and the CloudWatch setup used later in the post.
What the pilot actually checks
Before widening access, AWS suggests three checks. First, sign in with the IAM Identity Center profile through the AWS CLI and call Bedrock Runtime directly with a short prompt, which separates AWS access problems from Codex configuration problems. Second, point Codex at the approved model in its configuration file, setting the provider to amazon-bedrock, the Codex model ID, the profile and the Region. AWS advises keeping those settings in user-level or managed configuration so a repository cannot change the provider. Third, run a safe read-only Codex task against a non-production repository.
Prerequisites include a dedicated group for pilot users only, model inference permissions limited to what Codex needs plus the chosen observability destination, and a named rollback owner. Verification covers the assumed-role identity, model access through the selected Region, and completion of the read-only task.
Limits and open questions
The post promises to cover access, governance and cost decisions, but the excerpt published here gives no pricing, token rates or budget numbers. It also does not state which Regions carry which models, whether availability differs by inference path, or what licence or subscription Codex itself requires. Wider rollout steps, telemetry details beyond CloudWatch and the allowed-skills and MCP server policy are described only as things to decide, not as prescribed settings. Anyone planning a production rollout will need those specifics from the repository and the Bedrock model documentation.
What it means for people building apps
The guide is aimed at enterprises standardising coding agents on infrastructure they already govern, and its useful idea is separating identity, gateway and telemetry decisions rather than treating sign-in as the whole rollout. For smaller teams building apps with AI, most of that apparatus is unnecessary; the equivalent decision is simply which model an app's assistant calls. Draftship's notes on using the AI chat and choosing a model cover that choice at a much smaller scale. Draftship is not affiliated with AWS or OpenAI, and this announcement does not describe Draftship support for Bedrock.
Primary sources
- OpenAI Codex on Amazon Bedrock · AWS Machine Learning Blog, Nov 11, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.