Amazon Bedrock customers can now deploy OpenAI’s most advanced models directly from their AWS console. On June 1, 2026, the cloud giant quietly added GPT-5.5, GPT-5.4, and Codex to its fully managed AI service, triggering a wave of interest among enterprise architects who want the best of both clouds without the headaches of self-managed infrastructure.

This move marks the first time these frontier OpenAI models are available outside Microsoft Azure, shattering the longstanding exclusivity that tied OpenAI’s largest language models to a single hyperscaler. For enterprises running Windows workloads on AWS, the integration unlocks new possibilities around coding, data analysis, and governance that previously required complex cross-cloud plumbing.

What landed on Bedrock?

Bedrock’s model catalog now lists three new entrants under the OpenAI provider banner:

  • GPT-5.4: A multimodal model with a 500K token context window, designed for document understanding, financial analysis, and building retrieval‑augmented generation (RAG) pipelines with minimal hallucination. AWS has tuned its native Bedrock Knowledge Bases to work natively with GPT-5.4’s JSON‑native function‑calling syntax.
  • GPT-5.5: The bigger reasoning engine, boasting chain‑of‑thought depth that rivals o3‑class models. Early benchmarks show a 12‑point improvement on GPQA Diamond over GPT‑5.4, positioning it for regulated industries where audit trails must be reproducible.
  • Codex (v3): The code‑generation workhorse optimized for real‑time agentic workflows. It supports 24 programming languages out of the box and integrates with AWS CodeWhisperer custom actions, letting development teams scaffold entire serverless backends from natural‑language prompts.

All three models operate under Bedrock’s existing on‑demand and provisioned throughput pricing, with per‑token costs roughly 8–15% higher than equivalent Azure OpenAI Service SKUs—a premium that AWS argues is offset by zero data‑egress charges within the Bedrock ecosystem and the ability to keep everything inside a private VPC.

The cross‑cloud reality

For two years, enterprises that standardized on AWS for compute and storage had two messy options when they wanted GPT‑class models: either spin up an Azure subscription and duct‑tape cross‑cloud networking together, or settle for open‑source alternatives that lagged in reasoning quality. Bedrock’s onboarding of GPT‑5.4 and friends eliminates that friction.

“We run our ERP on RDS, our container fleet on EKS, and our CI/CD pipelines on CodePipeline. Having GPT‑5.5 in the same region as our database means we can finally build a real‑time fraud‑detection layer that doesn’t violate our data‑residency obligations,” said Marcus Liu, CTO at a Frankfurt‑based insurtech that beta‑tested the integration for six weeks. Liu’s team trained a custom Bedrock guardrail that redacts German insurance IDs before prompts ever hit the model, a compliance trick that would have been impractical when data crossed cloud boundaries.

AWS has also made multi‑cloud stinginess a feature, not a bug. Bedrock’s model‑routing API now lets you create “model groups” that mix OpenAI, Anthropic, and Mistral endpoints under a single IAM policy. An application can call GPT‑5.5 for complex reasoning, fall back to Claude Opus‑5 for contracts, and still log every token in CloudTrail as if it were a native AWS service.

Governance layers: what Windows admins need to know

Windows enterprise shops—especially those running Active Directory, Group Policy, and Intune hybrid‑joined devices—often struggle to enforce consistency when AI tools multiply. Bedrock’s integration with AWS IAM Identity Center now supports SAML 2.0 federation from Entra ID (formerly Azure AD), meaning a Windows domain user who signs into the AWS console via SSO automatically inherits the permission set their cloud architect defined.

More critically, AWS extended its bedrock:ModelInvocationLogging feature to the OpenAI models. Every prompt and response—including streaming chunks—drops into CloudWatch Logs with a vector‑clock‑stamped request ID that corrates with Bedrock’s guardrail decisions. Security teams can write AWS Config rules that automatically quarantine a model if it generates content that trips a specified word‑filter policy, all without touching the application code.

“We have 4,000 Windows endpoints running legacy .NET apps that we can’t recompile overnight,” said Elena Vasquez, infrastructure lead at a Tier‑1 North American bank. “With Bedrock, I can put an API proxy between the app and GPT‑5.4, enforce our data‑classification tags through API Gateway, and have CloudTrail prove to auditors that no PII left the VPC. That’s the kind of control that keeps regulators quiet.”

Codex and the developer workflow upgrade

The real sleeper announcement might be Codex. While GPT‑5.5 grabs headlines for reasoning, Codex slips neatly into the AWS development toolchain. When paired with Amazon Q Developer, Codex can build an entire CRUD microservice—writing a CDK TypeScript stack, generating Lambda handlers, and spitting out unit tests—from a single English description. The generated code follows the organization’s bedore‑defined repository structure because Q Developer passes the workspace context as a system prompt prefix.

Early adopters report 40% faster time‑to‑merge on feature branches when Codex participates in pull‑request reviews, flagging not just syntax errors but logic that contradicts the acceptance criteria spelled out in the PR description. The model also understands the nuances of the Windows‑centric .NET 11 runtime, generating source‑generators for minimal APIs that take advantage of native AOT compilation on AWS Graviton4 instances.

What’s missing and what’s next

No launch is flawless. Bedrock’s implementation of OpenAI models does not yet support streaming tool‑use—the most requested feature from the preview group—though AWS’s launch page promises an update “within the second half of 2026.” Additionally, fine‑tuning is unavailable; customers who need domain‑specific customization must lean on Bedrock’s retrieval‑augmented generation pipelines or wait for a future release.

Pricing, while competitive, still raises eyebrows. A single GPT‑5.5 inference costs roughly $0.03 per thousand input tokens, double Claude Opus‑5’s rate. But for workloads where reasoning quality is paramount—legal contract analysis, FDA‑submission drafting, or multi‑step mathematical proofs—the premium may be justifiable.

Looking ahead, AWS product managers have publicly hinted at a “BYO model weights” program that would let organizations host custom‑tuned versions of GPT‑5.x on Amazon S3 and deploy them through Bedrock’s private model feature. That would effectively turn Bedrock into a self‑service MLOps control plane for proprietary models, a direct counter to Azure’s Foundry platform.

What it means for Windows shops

If you manage a Windows‑centric enterprise on AWS, the calculus is now simpler. You can keep your flagship AI models in the same virtual neighborhood as your SQL Server Always‑On clusters, your Managed Active Directory, and your .NET ASP.NET Core APIs—all governed by the same set of IAM policies, KMS keys, and CloudTrail trails. The multi‑cloud tax just shrank substantially.

The next six months will reveal whether the cost premium wins over the convenience, but one thing is undeniable: OpenAI’s models are no longer an Azure exclusive. For Windows shops that bet big on AWS infrastructure, that changes the AI strategy immediately.