DEV Community

News
News

Posted on

AWS publishes official enterprise guidance for OpenAI Codex on Amazon Bedrock with GPT-5.6 Sol, Terra and Luna as inference back

Infographic cover

๐Ÿ“… November 11, 2026 โ€ข Source: AWS Machine Learning Blog

AWS published OpenAI Codex on Amazon Bedrock, official guidance for enterprises that want OpenAI Codex to use Amazon Bedrock for inference with supported OpenAI models. The models named in the post are GPT-5.6 Sol, Terra and Luna, with per-region availability checked in the Bedrock docs. Codex itself still runs on the developer workstation โ€” the main design choice is how its model requests reach Amazon Bedrock and where the organization applies policy and operational controls.

The guidance lays out three rollout paths. Native AWS Access uses workforce identity through IAM Identity Center โ€” for example Okta SAML sign-in, SCIM provisioning and temporary AWS credentials โ€” and starts with a least-privilege pilot. The Amazon Bedrock AgentCore Gateway path adds AWS managed gateway infrastructure for routing plus Amazon Bedrock Guardrails. The third path routes requests through an existing LLM Gateway to enforce controls in the request path, and a custom integration is described as the advanced form of that pattern, not a quick start. The companion open-source repository, Guidance for Codex on AWS at github.com/openai-on-aws/guidance-codex, ships quick starts for Native AWS Access, AgentCore Gateway, LiteLLM on ECS, Portkey, AWS Distro for OpenTelemetry (ADOT) and the Amazon CloudWatch setup.

Validation practices are concrete: a two-user identity isolation check confirms that assumed-role sessions, audit records and telemetry remain distinct, and rollback verification confirms that a new invocation is denied after access is removed. Rollout runs in phases from preparation through pilot and expansion to organization-wide standardization, with an administrator launch checklist. For observability, Codex OpenTelemetry metrics โ€” token usage, turns, API requests and tool calls โ€” flow through ADOT into CloudWatch dashboards, alongside Amazon Bedrock service metrics, model invocation logs and the AWS Cost and Usage Report for billed-usage evidence. The guidance says not to present token volume as a cost report and requires developers to review and test generated changes before merge. Model provider, credentials and Region stay in user-level or managed configuration, while repository-local .codex/config.toml and AGENTS.md are durable guidance, not enterprise enforcement boundaries.

Key figures

IMPACT. In practice, platform teams add a governed coding agent without building a second model-access, identity and operations stack. Security teams get IAM policies and audit trails, finance teams get billed-usage evidence, and MCP servers are treated as separate trust boundaries. An enterprise can pilot Codex on Bedrock with workforce identity in a least-privilege scope and then expand it under existing governance tooling as usage grows.

Sources: aws.amazon.com - github.com

IMPACT checklist


Sources:


Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky

AI Automation

Top comments (0)