Amazon Bedrock GPT-5.5: OpenAI Codex Routing, AWS IAM Credentials, and Model IDs

OpenAI's frontier models—GPT-5.5 and GPT-5.4—and the Codex coding agent are now generally available on Amazon Bedrock. For enterprise teams evaluating this path, the headline is less about model access and more about a credential architecture that is architecturally incompatible with the way most OpenAI API gateways work today.

What Actually Changed

Until now, every OpenAI request followed a single credential plane: OPENAI_API_KEY → api.openai.com. Routing layers, gateways, and proxies were built around this assumption.

The Bedrock path replaces all of that:

• Authentication: AWS IAM credentials or a Bedrock API key (AWS_BEARER_TOKEN_BEDROCK), not OPENAI_API_KEY.
• Base URL: Amazon Bedrock's regional endpoint, not api.openai.com.
• Model IDs: openai.gpt-5.5 and openai.gpt-5.4—AWS-namespaced strings, not OpenAI's own model names.
• API protocol: Bedrock's own implementation of the OpenAI Responses API—similar interface, but Bedrock is the request handler.
• Region dependency: Model availability varies by AWS region. You must select a region where the model is active.

For Codex specifically, the configuration change is explicit: model_provider = "amazon-bedrock" in ~/.codex/config.toml replaces any existing OpenAI provider config.

The Routing Implication: Two Credential Planes, Not One

Most OpenAI-compatible gateways today assume a homogeneous credential set—one API key per provider, one base URL per provider. The Bedrock path breaks that assumption in several ways:

1. You cannot reuse OPENAI_API_KEY. Bedrock authentication is AWS-native. If your routing layer injects Authorization: Bearer <openai-key> before forwarding to Bedrock, the request will fail. The credential must be resolved against AWS's credential chain (env vars, shared credentials file, IAM instance profile, SSO, or a federated credential_process).

2. Model ID namespace is different. A request for gpt-5.5 sent to Bedrock won't resolve. The correct ID on Bedrock is openai.gpt-5.5. Any routing policy that maps logical model aliases to provider-specific IDs needs a Bedrock-aware alias table.

3. Regional routing adds a new variable. Unlike the global api.openai.com, Bedrock endpoints are region-scoped (e.g., us-east-1, eu-central-1). Your routing policy needs to know which region a team is provisioned in to target the right endpoint—and which models are active there.

4. Pricing matches OpenAI first-party rates. This removes the pricing-arbitrage rationale for routing to Bedrock over direct OpenAI. The decision is driven by governance and compliance, not cost.

What's Not Available on Bedrock (Yet)

Before routing traffic to Bedrock, teams should audit against the current capability surface:

Fast Mode is a significant gap for latency-sensitive coding agent workflows. Cloud agents—Codex review, security scan, web agent—remain OpenAI-hosted-only for now.

What Operators Should Do

Audit your credential injection logic. If your gateway unconditionally injects OPENAI_API_KEY on all outbound requests, add a Bedrock-aware path that resolves AWS credentials through the SDK chain instead.

Add Bedrock model aliases. Update your model alias table to map gpt-5.5 → openai.gpt-5.5 when the target provider is Bedrock. Do not send OpenAI-namespace model IDs to Bedrock endpoints.

Pin routing per team. Since Bedrock availability varies by region and not all orgs have the same provisioned access, routing to Bedrock should be an explicit per-tenant or per-workspace policy, not a global default.

Check feature parity before migrating Codex workloads. If your team uses cloud agents, Fast Mode, or hosted plugins, migrating Codex to the Bedrock path today will silently degrade those capabilities.

Track the Daybreak roadmap. OpenAI has indicated that Daybreak—their cyber defense model and Codex Security offering—will follow to Bedrock. Security teams on AWS may want to watch this path before building their secure code review pipeline.

The Governance Upside

The reason enterprises will use this path is clear: procurement through existing AWS accounts, data residency within a selected Bedrock region, unified billing, no per-seat licenses, and AI governance controls already embedded in the AWS organization's guardrails. For regulated industries—healthcare, finance, government—this may be the only viable path to deploying GPT-5.5 in production.

This is the same pattern we saw when Claude arrived on Bedrock and Vertex: the model is the same, but the operational context—who owns the key, who sees the request, which compliance attestations apply—changes completely.

Related

• Claude Code Auto-Mode on Bedrock, Vertex, and Foundry
• OpenAI's Workload Identity and Admin API Governance Changes
• OpenAI's Codex Goes On-Premises: The Dell Partnership