OpenAI’s GPT-5.5 and Codex Land on AWS Bedrock: Frontier Models Inside Your AWS Account

For a long time, adopting a frontier model and keeping your security team happy felt like two competing goals. The best models lived behind one vendor’s API; your data, governance and billing lived somewhere else. AWS has just narrowed that gap. OpenAI’s newest models and its coding agent are now generally available on Amazon Bedrock, which means you can use them for real production workloads without leaving the account, the controls and the bill you already run everything else on.

What AWS announced

OpenAI’s GPT-5.5 and GPT-5.4 are now generally available on Amazon Bedrock, accessed through the Responses API, alongside Codex, OpenAI’s coding agent. “Generally available” is the part that matters: this is no longer a preview to experiment with, but a supported path for production use.

Two details stand out. First, the commercial terms are straightforward: you pay the same per-token rate as you would direct from OpenAI, with no additional fees for running them through Bedrock. Second, Codex is not locked to a single surface. It is available through the Codex app, the Codex CLI and IDE integrations for Visual Studio Code, JetBrains and Xcode, so it can meet developers in the tools they already use. AWS notes that more than five million people already use Codex every week for software development.

Why running inside Bedrock matters

The headline is the model; the substance is where it runs. Because these models sit inside Bedrock, they inherit the AWS controls your organisation has already adopted, reviewed and audited.

• Access control: IAM permissions govern who and what may call the models, using the same policies you apply across the rest of your estate.
• Network isolation: VPC and PrivateLink keep traffic off the public internet where that is a requirement.
• Encryption: KMS encryption protects data with keys you manage.
• Auditability: CloudTrail logging records who called what and when, which is the evidence a compliance review actually asks for.

AWS also states that your prompts and responses are not used to train models. Taken together, these are the practical answers to the questions that usually stall an AI project: who can use it, where the data goes, how it is protected, and how you prove all of that after the fact.

What it means for businesses

For organisations already on AWS, this removes a familiar source of friction. Adopting a frontier model no longer means onboarding a new vendor, negotiating a separate data-processing agreement, threading another set of credentials through your systems, or reconciling a second invoice. The model becomes one more capability inside an account your security and finance teams already understand.

The use cases AWS highlights are squarely about getting work done: writing and debugging code across large codebases, refactoring, testing and validation, analysing data, generating documents and spreadsheets, and automating multi-step tasks. In other words, this is aimed at agentic coding and knowledge work: the kind of repetitive, rules-bound effort that is expensive to do by hand and well suited to careful automation.

What to do about it

Availability is an opportunity, not a strategy. The temptation is to switch a model on and call it adoption; the better move is to pick one well-scoped job and do it properly. Start where the value is concrete and the risk is contained: a developer-automation workflow, a document-generation task, a data-analysis step that a person currently repeats by hand.

• Scope tightly: choose one task with a measurable outcome rather than a vague plan to “add AI”.
• Use your controls: set IAM permissions, network isolation and key management from the start, so governance is built in rather than retrofitted.
• Watch the cost: per-token pricing is predictable, but agentic workflows make many calls; measure spend against the value of the work.
• Keep a human in the loop: let the model prepare consequential work and a person approve it before it lands.

Do that, and the model on Bedrock stops being a line item to justify and becomes a dependable part of how the work gets done.

Source: AWS: OpenAI models and Codex on Amazon Bedrock are now generally available