Codex can now run on Amazon Bedrock, with AWS logins instead of OpenAI keys

Amazon Web Services has published a guide to running OpenAI’s Codex on Amazon Bedrock. Codex stays on the developer’s machine, while its model requests go to Bedrock through AWS identity and IAM controls, which platform teams already operate.

Three ways to route Codex through AWS

AWS describes three access paths. Native AWS Access is the recommended default for a small, least-privilege pilot. Staff sign in through their existing identity provider, which feeds AWS IAM Identity Center, and that issues the temporary AWS credentials Codex uses. The post’s example uses Okta with SAML and SCIM, and says the identity provider is not configured in Codex itself.

The second path is AgentCore Gateway, part of Amazon Bedrock AgentCore, for teams that want AWS-managed routing or Amazon Bedrock Guardrails. The third is an LLM gateway, meant for organizations that already run one or want to enforce controls in the request path. A fully custom integration is described as the advanced form of that pattern, not a quick start. The examples come from an open source repository, Guidance for Codex on AWS, which includes quick starts for native access, AgentCore Gateway, LiteLLM on ECS and Portkey.

The pilot has three checks. First, sign in with aws sso login and call Bedrock Runtime directly with aws bedrock-runtime converse to rule out AWS-side problems. Then set the provider in ~/.codex/config.toml: model_provider = "amazon-bedrock", the model openai.gpt-5.6-sol, plus an AWS profile and Region. Finally, run a safe task in a non-production repository. The example uses us-east-2. AWS tells readers to confirm in its supported-models table whether Sol, Terra or Luna is available in the Region and inference path they plan to use.

Attribution and cost need more than Codex telemetry

The post separates what each data source can prove. Codex’s OpenTelemetry metrics, collected through an ADOT collector and shown in a CloudWatch dashboard, describe client activity such as tokens, turns and tool calls. Per-user grouping works only after a local collector adds validated identity attributes. Who invoked the model comes from AWS identity and audit records, and actual cost from billing data. For native IAM Identity Center access, AWS says Cost and Usage Report 2.0 with caller-identity allocation can expose the SSO principal. Gateway paths need their own telemetry and spend records.

AWS also asks for a two-user check, so that separate sessions and audit records stay distinct, and a rollback test: remove the configuration or assignment, then verify that a new invocation is denied. Provider, credential and telemetry settings belong in user-level or managed configuration, not in a project-level .codex/config.toml, which AWS says is not an enforcement boundary. The same goes for AGENTS.md.

What to watch as Codex spreads inside AWS accounts

The guide lands in a week of OpenAI and AWS tie-ups. According to our earlier report, OpenAI listed ChatGPT Enterprise in AWS Marketplace on November 9, covering ChatGPT, ChatGPT Work and Codex, and Amazon says GPT-5.6 pricing on Bedrock matches OpenAI’s first-party rates and counts toward AWS commitments. That article also lists the Bedrock timeline: limited preview on April 28, general availability of GPT-5.6 Sol, Terra and Luna on July 9, and GPT-6 Astra on September 8.

On November 10, AWS also added bedrock-runtime endpoint support for OpenAI’s Daybreak security models, GPT-5.6 Sol among them as Daybreak Blue, through US and Global cross-Region inference profiles. Together, these show AWS making Bedrock a place where OpenAI’s coding and security tools can be bought and governed like other cloud services.

The post does not establish several things a buyer would want. It gives no Codex pricing on Bedrock, no Region list, and no performance or productivity figures. It also publishes no test results from customers. Its advice is operational and comes from AWS itself, and the claims about the open source repository’s configurations have not been independently checked. Signals worth tracking are which Regions carry Sol, Terra and Luna, whether AgentCore Gateway gains Codex-specific controls, and whether enterprises that already hold AWS commitments choose Bedrock over a direct OpenAI contract for Codex.