Codex on Amazon Bedrock: AWS Setup Guide

Amazon Bedrock offers GPT-6 Astra alongside GPT-5.6 Sol, Terra, and Luna. AWS model availability and Codex client support are separate checks. This guide configures Sol through the built-in amazon-bedrock provider and compares Astra’s model IDs, regional routes, and costs before a team rollout.

Status by capability

Surfaces and feature limits

Codex on Bedrock is a strong path for local engineering workflows. It does not turn every OpenAI-hosted feature into a Bedrock feature.

AWS account strategy for pilots

Start with an isolated account or sandbox OU. The AWS account boundary remains one of the simplest ways to control billing, permissions, CloudTrail, Service Control Policies, budgets, and blast radius.

Mantle inference requires bedrock-mantle:CreateInference. Runtime uses bedrock:InvokeModel and streaming permissions, with Responses access also covering the default project.

Region, residency, US and Canada

For this built-in Codex Mantle setup, Sol is documented in us-east-1 and us-east-2; Terra and Luna also support commercial us-west-2. The Codex guide excludes GovCloud Mantle endpoints. These are in-Region routes. Runtime uses different cross-Region profile IDs: its source Region does not guarantee where inference runs. Check the model’s endpoint and Region table before making residency commitments. For Astra, Mantle inference runs only in us-west-2 (Oregon). Runtime requires a US or global cross-Region profile; it does not offer in-Region Astra inference. Moving from this Sol example to Astra therefore changes both model and endpoint planning.

For US teams, choose the region by model availability, latency, compliance, and operational proximity. For Canadian organizations, document whether a pilot may use a US region with low-risk data or must wait for an approved region. Keep sensitive data out of the pilot until the required model, endpoint, quota, and network path are approved.

Configure Codex with Amazon Bedrock

The official Codex Bedrock guide describes the amazon-bedrock provider using Bedrock Mantle. The local client sends model requests to Bedrock’s OpenAI-compatible Responses API; the OpenAI-hosted Responses API is not in the request path. Authentication uses a Bedrock API key or the AWS SDK credential chain.

AWS recommends Runtime for most new API applications; that recommendation does not change this built-in Codex configuration.

The Codex Bedrock support list contains openai.gpt-5.6-sol, openai.gpt-5.6-terra, openai.gpt-5.6-luna, openai.gpt-5.5, and openai.gpt-5.4. Use Sol for demanding coding and reasoning, Terra for balanced everyday work, and Luna for lower-cost, high-volume tasks. The AWS cards for Sol, Terra, and Luna list text/image input, text output, and 1M context; OpenAI’s API guide specifies 1,050,000 tokens. Confirm the deployed limit and budget for the higher price tier above 272K input tokens.

Set the provider and model in ~/.codex/config.toml:

model = "openai.gpt-5.6-sol"
model_provider = "amazon-bedrock"

Authentication: Bedrock API key or AWS SDK

Codex checks first for a Bedrock API key, then the AWS SDK credential chain. For API-key authentication, set the region explicitly:

export AWS_BEARER_TOKEN_BEDROCK="REPLACE_WITH_YOUR_BEDROCK_API_KEY"
export AWS_REGION=us-east-2

For organizations using SSO, prefer temporary credentials through an AWS profile:

aws configure sso --profile codex-bedrock
aws sso login --profile codex-bedrock
export AWS_PROFILE=codex-bedrock
export AWS_REGION=us-east-2

Desktop apps and IDE extensions may not inherit shell environment variables. For those surfaces, place required values in ~/.codex/.env and restart the app or extension.

Troubleshoot authentication and verify the provider

Codex checks AWS_BEARER_TOKEN_BEDROCK before the AWS SDK credential chain. When testing SSO, remove any unintended Bedrock API key from the environment read by the client, including ~/.codex/.env. Restart desktop or IDE sessions, then use /status in the CLI to confirm the amazon-bedrock provider.

If authentication fails, check the exact model ID, AWS Region, model permissions, and credential expiry. For SSO, renew the selected profile with aws sso login --profile codex-bedrock. A successful SSO login does not override an API key still present in the client environment.

Cost, audit, and private networking

Pricing and budgets

According to the June 1, 2026 AWS announcement, GPT-5.5 and GPT-5.4 launched with access through the Responses API on Bedrock at the same per-token rate as direct OpenAI usage with no additional fees. Codex on Bedrock also uses pay-per-token pricing, with no seat licenses or per-developer commitments, and usage counts toward existing AWS commitments.

Current price parity applies to equivalent services. OpenAI’s Bedrock guide equates regional Bedrock processing with OpenAI regional processing, which carries a 10% uplift over global API rates. Sol’s AWS card lists short-context in-Region prices of $4.40 input, $5.50 cache write, $0.44 cache read, and $22 output per million tokens. Direct OpenAI global prices are $4, $5, $0.40, and $20 respectively. Match context tier, routing, and cache usage before comparing costs. OpenAI says Sol’s promotional pricing lasts at least through November 21, 2026.

For an Astra evaluation, use the AWS Standard rates for the selected route and context tier. The regional prices below are per million tokens:

Astra’s context window is 1,050,000 tokens, with up to 128,000 output tokens. Its short-context regional rates are 2.5 times Sol’s for the same token mix; that is a price comparison, not a measurement of cost per completed task. Benchmark completion, retries, and token usage before choosing it for a workflow.

• Cost: Codex on Bedrock is billed per token, not per seat. Create an account budget, an Amazon Bedrock service-filtered budget, and shared alerts before rollout.
• Attribution: use separate accounts, tags, Projects, and cost reports to separate experiment, development, and production usage.

Audit and private networking

The built-in Codex provider uses Mantle. Bedrock invocation logging covers Runtime, including Responses and Chat Completions there, but not Mantle. Keep gateway records for Mantle when request/response capture is required.

• Audit: enable centralized CloudTrail and review Bedrock events from day one.
• Network: evaluate VPC endpoints and PrivateLink for bedrock, bedrock-runtime, and bedrock-mantle when traffic must stay private.
• Data: start with low-risk repositories and tasks before allowing personal, customer, or production data.

Retention requirements

For GPT-5.6 Sol, AWS may retain classifier-flagged traffic for up to 30 days; eligible customers can request full ZDR through their AWS account team. default and Responses store=false do not guarantee ZDR. Check the selected model’s retention requirements and available settings before rollout. Customer-managed gateway and invocation logs need separate retention controls. See the AWS abuse-detection documentation.

Rollout checklist

• Exact supported model ID selected and regional availability validated; this guide uses openai.gpt-5.6-sol in us-east-2.
• Codex updated and amazon-bedrock provider confirmed in /status.
• Credentials selected: Bedrock API key or AWS SSO/AWS profile with temporary credentials.
• AWS account or sandbox OU created with budgets, alerts, and CloudTrail.
• Repository policy, sensitive-data policy, human approval, and tool-use policy approved.
• PrivateLink, endpoint policies, SCPs, and KMS evaluated based on risk.
• First task run against a low-risk repository, with cost and logs reviewed.
• Operational owner, rollback, and support process defined before expanding to more developers.

FAQ

Does Codex on Bedrock use ChatGPT login?

No. With the amazon-bedrock provider, Codex uses AWS-native authentication: Bedrock API key or the AWS SDK credential chain. Do not use ChatGPT sign-in or OPENAI_API_KEY for this path.

Is gpt-oss the same as GPT-5.6?

No. gpt-oss is a separate open-weight family with different IDs, capabilities, and prices.

Do we need PrivateLink?

Not every pilot needs it, but regulated environments should evaluate PrivateLink and endpoint policies early, especially when prompts include code, personal data, or customer data.

How Elevata helps

This work takes more than a config.toml file. Elevata helps platform, security, and engineering teams validate OpenAI on Amazon Bedrock: account and OU design, IAM/SCP, Codex rollout, Mantle vs Runtime, PrivateLink, CloudTrail, Budgets, Projects, observability, residency, and developer governance.

If you want to release Codex, GPT-5.6, or OpenAI workflows on AWS without losing control of cost and security, validate AWS readiness with Elevata.