OpenAI on Amazon Bedrock: Models, Pricing, and How It Works in 2026

OpenAI on Amazon Bedrock

For years, if you ran your stack on AWS and wanted OpenAI's models, you had to route around them. OpenAI's frontier models lived on OpenAI's own platform and on Microsoft Azure, not on Amazon Bedrock. That changed in 2026. OpenAI on Amazon Bedrock is now a first-class option: GPT-5.5, GPT-5.4, and the Codex coding agent are generally available, joining the open-weight gpt-oss models that arrived in 2025.

For AWS customers this is a meaningful shift. You can now build with OpenAI's most capable models using the same security, identity, billing, and governance you already use across AWS, and at the same per-token price OpenAI charges directly, with no AWS markup. This guide covers what is available, what it costs, how access works, why it happened now, and how OpenAI on Amazon Bedrock compares to using OpenAI directly or through Azure.

Key takeaway: OpenAI on Amazon Bedrock means GPT-5.5, GPT-5.4, Codex, and the open-weight gpt-oss models run on AWS infrastructure through the Responses API. Pricing matches OpenAI's first-party rates with no additional fees, and usage counts toward your existing AWS commitments. You get OpenAI model behavior wrapped in AWS identity, VPC isolation, encryption, and billing. The initial launch is US-region first, so check model and Region support before rollout.

What OpenAI on Amazon Bedrock Actually Means

Amazon Bedrock is AWS's managed platform for building and running AI applications and agents at production scale. It already hosts models from Anthropic, Meta, and others behind a single API. The 2026 expansion adds OpenAI's frontier models to that catalog, so a team standardized on Bedrock no longer has to leave AWS to use GPT.

The timing was not accidental. OpenAI and Microsoft restructured their partnership in late April 2026, ending Microsoft's exclusive cloud arrangement. Within a day, OpenAI announced availability on AWS. A limited preview followed almost immediately, and roughly one month later, on June 1, 2026, GPT-5.5, GPT-5.4, and Codex reached general availability on Amazon Bedrock. The open-weight gpt-oss-120b and gpt-oss-20b models had already been available on Bedrock and SageMaker since August 2025, marking the first time any OpenAI models ran on Amazon's cloud.

Which OpenAI Models Are on Amazon Bedrock?

Five OpenAI options are available, spanning frontier reasoning, a coding agent, and low-cost open-weight models.

GPT-5.5 and GPT-5.4 are called through the Responses API and run on Bedrock's next-generation inference engine. Codex routes its inference through Bedrock while you keep using the Codex App, the Codex CLI, or IDE integrations for Visual Studio Code, JetBrains, and Xcode.

OpenAI on Amazon Bedrock Pricing

This is the headline most AWS teams care about: pricing matches OpenAI's first-party per-token rates, with no additional Bedrock fee on top, and usage counts toward your existing AWS commitments. That last point matters because it means OpenAI spend can draw down the same commitments you already negotiated with AWS.

No markup is the real story

Unlike some managed routes that layer infrastructure overhead on top of token costs, OpenAI on Amazon Bedrock charges the same rate as OpenAI direct with no additional fees. You are paying for the models, not for the privilege of running them on AWS, while still getting AWS security and billing. Note that global cross-region inference pricing for OpenAI models is still rolling out.

For the wider Bedrock cost picture across all hosted models, our Amazon Bedrock pricing guide breaks down deployment modes and fees, and our ChatGPT pricing in 2026 guide covers the same OpenAI rates from the first-party side.

How Access and Inference Work

OpenAI models on Bedrock run through an AWS-managed deployment path that is compatible with the OpenAI Responses API. Your application keeps using OpenAI model behavior, but AWS owns the surrounding control plane: account access, Regional availability, identity, and billing.

• Responses API and model IDs. You call models with IDs like openai.gpt-5.5, openai.gpt-5.4, openai.gpt-oss-120b, and openai.gpt-oss-20b, using the OpenAI SDK pointed at Bedrock endpoints or a BedrockOpenAI client.
• AWS-native security. Authentication uses AWS IAM, with VPC isolation and encryption, the same controls you already apply to other AWS workloads.
• Resilient inference. Each request gets an isolated queue with automated capacity management. Request state is captured durably, so if a node restarts mid-call the request resumes instead of starting over. During high demand, requests are queued rather than rejected.
• Region-first rollout. At general availability, GPT-5.5 launched in US East (Ohio) and GPT-5.4 in US East (Ohio) and US West (Oregon). The initial scope is narrower than the full OpenAI API, so confirm model and Region support before you build.

Why This Matters for AWS Teams

The practical win is consolidation. Teams that already run Anthropic's Claude and other models on Bedrock can now add OpenAI to the same platform, the same API surface, the same governance, and the same invoice. There is no second vendor relationship to manage and no separate security review for a new cloud.

It also reshapes a common architecture decision. Previously, choosing OpenAI often meant accepting Azure or a direct OpenAI integration outside your AWS estate. Now the multi-model strategy lives in one place. If you are weighing OpenAI against Anthropic for a workload, our Anthropic vs OpenAI comparison lays out the tradeoffs, and both now sit side by side in the Bedrock catalog.

OpenAI on Bedrock vs OpenAI Direct vs Azure OpenAI

Since token rates are the same across all three, the decision comes down to governance, billing, and feature timing.

The contrast with Azure is striking. Azure OpenAI keeps the same token rate but adds 15 to 40 percent in support, networking, and infrastructure overhead, as we detail in our AWS vs Azure pricing guide. Bedrock's no-markup approach removes that premium for AWS-native teams. Use OpenAI direct when you need the very latest first-party features that the managed routes have not yet picked up.

Codex on Amazon Bedrock

Codex, OpenAI's coding agent, is one of the more interesting parts of this launch for engineering teams. You keep your existing Codex workflow through the App, CLI, and IDE plugins, but configure it to route inference through Bedrock, so the code your agent reads and writes stays inside your AWS security boundary with IAM, VPC isolation, and encryption. Pricing is pay-per-token at the underlying model rate, and that usage counts toward your AWS commitments like everything else.

If you are evaluating Codex cost and where it fits, our OpenAI Codex explained guide walks through the product, its plans, and how to keep its spend under control.

How to Keep OpenAI on Bedrock Costs Under Control

No markup does not mean no discipline. Token spend still grows fast, and because it now flows through AWS, it can hide inside a larger cloud bill.

• Route by difficulty. Send routine work to GPT-5.4 or the gpt-oss models and reserve GPT-5.5 for the hardest tasks. Output rates differ enough that this is the biggest lever.
• Tag and allocate. Because OpenAI usage counts toward AWS commitments, fold it into the same cost-allocation model as the rest of your AWS spend so you can see cost per team and per feature.
• Apply AI FinOps. Treat token spend like any other cloud cost, with budgets, anomaly alerts, and unit economics. Our FinOps for AI token and GPU costs guide and our LLM cost optimization guide cover the playbook, and the engineering side of attribution is in our cloud cost allocation guide.

Conclusion

OpenAI on Amazon Bedrock closes a long-standing gap for AWS customers. You get OpenAI's frontier models and Codex with AWS identity, isolation, encryption, and billing, at the same token rates OpenAI charges directly and with no AWS markup, while the spend draws down your existing AWS commitments. For teams already invested in Bedrock, it turns a multi-vendor headache into a single-platform choice between Anthropic, OpenAI, and open-weight models. The launch is US-region first, so confirm availability before you build, route work to the right model, and fold OpenAI spend into your AWS cost allocation from day one. If you want help attributing and governing OpenAI and the rest of your AWS spend, that is exactly the discipline FinOps brings.