GPT-5.5 on Amazon Bedrock: Complete Setup and Pricing Guide (2026)

Anastasiia Kuten Marketing & Partnerships Department Manager

June 8, 2026

AWS partner dedicated to startups

• 2000+ Clients
• 5+ Years of Experience
• $10M+ saved on AWS

OpenAI’s most advanced models are now running on AWS. Not through a third-party wrapper, not through a custom integration, but as first-class citizens inside Amazon Bedrock with the same APIs, IAM controls, and audit logging as every other Bedrock model.

GPT-5.5, GPT-5.4, and Codex hit general availability on Amazon Bedrock on June 1, 2026. This is the biggest enterprise AI story of the year, and most teams are still trying to figure out what it actually means for them. This guide covers everything: setup, pricing, regional availability, how it compares to Claude on Bedrock, and whether it is the right choice for your workload.

What Just Happened

For most of OpenAI’s history, the only way to use GPT models was through OpenAI’s API directly or through Microsoft Azure (under the exclusive cloud partnership that ended earlier this year). If you wanted GPT inside an AWS-governed environment, you had to wire up your own bridge between AWS and OpenAI, which most enterprises could not do without serious compliance headaches.

That ended in April 2026 when AWS and OpenAI announced an expanded partnership. The first batch (GPT-5.5, GPT-5.4, and Codex) went into limited preview at the end of April. On June 1, 2026, all three reached general availability on Amazon Bedrock.

You can now call OpenAI’s frontier models through the same Bedrock APIs you already use, with your existing IAM, KMS encryption, VPC isolation, CloudTrail logging, and AWS commitments. No new procurement, no separate vendor agreement, no infrastructure to build.

What Models Are Available

Pricing matches OpenAI first-party rates exactly. There are no additional AWS fees on top. Usage counts toward your existing AWS commitments, which means OpenAI spend can now apply against your Enterprise Discount Program (EDP), Savings Plans, and other AWS contracts.

What GPT-5.5 Is Actually Good At

OpenAI markets GPT-5.5 as their most advanced frontier model. In practice, the strengths that matter for enterprise use:

• Agentic coding: multi-step software development tasks, refactors, debugging, and long autonomous code generation. This is where Codex shines.
• Data analysis: reasoning over structured data, generating SQL, building analytical workflows.
• Multi-step autonomous tasks: agent workloads that run for hours, making decisions and using tools.
• Document and knowledge work: long-form analysis, summarization, research workflows.
• Tool use: reliable function calling, MCP integration, and API orchestration.

GPT-5.4 is described by AWS as the “best price-performance” model in the lineup. It is the model you reach for when GPT-5.5 is overkill and Codex is not the right fit.

How to Set It Up

Setup is mostly identical to enabling any other Bedrock model. Here is the practical walkthrough.

Step 1: Request Model Access

1. Sign in to the AWS Console and navigate to Amazon Bedrock
2. Open Model access in the left sidebar
3. Click Manage model access
4. Find OpenAI under model providers and select the models you want (GPT-5.5, GPT-5.4, and Codex)
5. Submit the request. Approval is typically instant for most accounts.

Model access is regional. If you need GPT-5.5 in eu-west-1, you have to request access in that specific region.

Step 2: Call the Model via the Responses API

This is the most important detail many teams miss: GPT-5.5 and GPT-5.4 on Bedrock only support the Responses API, not the older Messages or Converse APIs that other Bedrock models use.

The Responses API is OpenAI’s newer interface for agentic and tool-calling workloads. Your code will look familiar if you have used the OpenAI SDK directly:

from openai import OpenAI

client = OpenAI(
    base_url="https://bedrock-runtime.us-east-1.amazonaws.com/openai/v1",
    api_key="YOUR_BEDROCK_API_KEY"
)

response = client.responses.create(
    model="openai.gpt-5.5",
    input="Analyze this CSV and find the top 3 anomalies",
)

print(response.output_text)

If you have existing OpenAI-direct code, the migration is minimal: change the base URL and authentication, keep almost everything else the same.

Step 3: Configure Authentication

Two options:

Bedrock API Key: the simplest path. Generate a key in the AWS Console under Bedrock. Use it like any OpenAI API key. Good for development and quick prototyping.

AWS SDK Credentials: the production path. Uses the standard AWS SDK credential chain (environment variables, IAM roles, AWS SSO). Necessary for IAM-governed access, fine-grained permissions, and CloudTrail attribution.

Step 4: Set Up Cost Controls

This is the step most teams skip and regret. Before you put GPT-5.5 in production:

1. Create an AWS Budget filtered to Amazon Bedrock spend
2. Set alerts at 50%, 80%, and 100% of expected monthly spend
3. Enable AWS Cost Anomaly Detection on the Bedrock service
4. Tag your application with cost allocation tags (Environment, Project, Team)

GPT-5.5 at $30/M output tokens is one of the more expensive models on Bedrock. A runaway agent loop can rack up $10,000 in a weekend. Cost controls are not optional.

Step 5: Test with a Real Workload

Before migrating production traffic, run a representative workload through GPT-5.5 on Bedrock and compare results to your current setup (Claude on Bedrock, GPT-5.5 direct on OpenAI, whatever you have). Measure:

• Response quality on your actual tasks
• Latency at production load
• Token usage per request (impacts cost)
• Error rates and retry behavior

Bedrock provides its own isolated inference engine with automated capacity management. Performance should be excellent, but production-load testing always reveals things benchmarks do not.

Regional Availability

At GA on June 1, 2026, the OpenAI models on Bedrock are available in a subset of regions, with more added through 2026.

Confirmed regions at launch:

• US East (N. Virginia)
• US West (Oregon)
• Some EU regions (Ireland, Frankfurt)
• AWS GovCloud (US-West): GPT-5.4 added June 3, 2026

Bedrock offers three inference routing options:

For GDPR-bound workloads, use In-Region in an EU region. For maximum scale, use Global Cross-Region. For most production workloads, Geo Cross-Region is the right balance.

GPT-5.5 on Bedrock vs OpenAI Direct: Which Should You Use?

This is the question every team asks first.

The short answer: if you are already heavily on AWS and have compliance, procurement, or governance requirements, Bedrock is the obvious choice. If you are not on AWS, going through Bedrock just for OpenAI access does not make sense.

For most enterprise teams, Bedrock removes friction (compliance review, vendor onboarding, audit setup) that would have blocked OpenAI adoption before. That is the real value of this launch.

GPT-5.5 vs Claude Opus 4.8 on Bedrock

The other comparison everyone is asking about. With both OpenAI and Anthropic models on Bedrock, you can actually pick the right model per workload instead of being locked into one provider.

Both are excellent models. Test on your specific workload before committing. Some teams report Claude Opus 4.8 is stronger on long autonomous tasks. Others find GPT-5.5 stronger on data analysis and structured reasoning. Use Bedrock’s new console feature to compare them side by side on your actual prompts.

Codex on Bedrock: The Game-Changer for Coding Teams

Codex is worth its own section because it changes the cost structure for AI-assisted development entirely.

Before this launch, Codex was sold per-seat (you pay a subscription per developer per month). Codex on Bedrock drops seat licenses entirely and switches to pay-per-token billing.

For large teams, this is a meaningful change:

• Inactive developers cost nothing. A 500-person engineering org with 200 occasional Codex users no longer pays for 500 seats.
• Heavy users cost more. A developer running Codex agents 24/7 will spend more than they would on a seat license. Plan for usage-based budgeting.
• Usage counts toward AWS commitments. Codex spend reduces your AWS Enterprise Discount Program (EDP) burn-down obligations.

Codex on Bedrock works with the Codex App, the Codex CLI, and IDE integrations for VS Code, JetBrains, and Xcode. The Codex configuration uses the amazon-bedrock provider. Codex runs locally on your machine and sends model requests to Bedrock. Authentication uses either a Bedrock API key or the AWS SDK credential chain (Codex checks the API key first).

For Anthropic’s Claude-based coding agent, see our guide on Kiro, AWS’s spec-driven AI coding platform. Many teams use both: Codex for OpenAI-friendly workflows, Kiro for Claude-friendly workflows.

What This Means for Existing AWS Teams

Several things change for enterprises already on AWS.

Compliance barriers drop. Teams in regulated industries (healthcare, financial services, government) that were blocked from using OpenAI directly because of data residency, audit, or contracting issues can now adopt it inside their existing AWS governance.

Procurement simplifies. No new vendor onboarding, no separate billing, no contract negotiation. Spend rolls up under existing AWS agreements.

Model selection becomes a real choice. With Claude, Llama, Titan, Nova, Mistral, and now GPT-5.5 all on Bedrock, you can pick the best model per task instead of standardizing on one provider. The Bedrock console now supports side-by-side model comparison.

Multi-model strategies become easier. Route simple tasks to GPT-5.4 or Claude Haiku, complex tasks to GPT-5.5 or Claude Opus 4.8, coding tasks to Codex. All through one API surface.

Bedrock becomes the default AI platform. For most enterprise AWS customers, there is no longer a reason to integrate OpenAI directly. Bedrock is the right choice.

Limitations and Honest Concerns

This is the section AWS marketing will not write.

Responses API only. If your codebase is built on the Messages or Converse APIs, you need to add the Responses API path for OpenAI models. Not a big deal for new builds, real work for existing systems.

No fast mode. Unlike Claude Opus 4.8, GPT-5.5 does not have a fast/cheap tier. You pay full price for every call.

Console support still rolling out. At GA, full console support for GPT-5.5 and GPT-5.4 is “coming soon.” The Responses API works programmatically but the visual playground experience lags behind other Bedrock models.

Region availability gaps. If your data residency rules pin you to a specific region, verify model availability there. EU regions tend to lag US ones.

Managed Agents are still in limited preview. The companion product “Amazon Bedrock Managed Agents, powered by OpenAI” is not yet generally available. Plan it as a future option, not an immediate dependency.

Output token pricing is higher than Claude. $30/M output tokens vs $25 for Claude Opus 4.8. For output-heavy workloads, the cost difference adds up.

Should You Migrate Your OpenAI Workloads to Bedrock?

The decision framework.

Migrate now if:

• You are already heavily on AWS
• You have compliance or audit requirements that benefit from CloudTrail and KMS
• You need data residency in specific AWS regions
• You have an AWS Enterprise Discount Program or Savings Plan commitment to burn down
• You want centralized procurement and billing
• You are evaluating multiple models (Claude, GPT, Llama) and want one API surface

Stay on OpenAI direct if:

• You are not on AWS and have no plans to migrate
• You need OpenAI features that are not yet on Bedrock (latest beta features, specific fine-tuning options)
• Your workload is tiny enough that integration overhead is not worth it
• You are deeply committed to Azure or another cloud

For most enterprise AWS customers, the answer is “migrate when you have the engineering capacity for the small API change required.” The benefits compound over time.

Frequently Asked Questions

Is GPT-5.5 on Bedrock more expensive than OpenAI direct?
No. Pricing matches OpenAI’s first-party rates exactly with no additional AWS fees on top. You pay the same per-token rate.

Does using GPT-5.5 on Bedrock count toward AWS Savings Plans?
Yes. Usage counts toward existing AWS commitments including the Enterprise Discount Program (EDP) and Savings Plans.

Is my data used to train OpenAI models when I use Bedrock?
No. Your prompts and outputs through Bedrock are not used to train any model. This is an architectural guarantee.

Can I use the existing OpenAI SDK with GPT-5.5 on Bedrock?
Yes, with minor changes. Change the base URL to your Bedrock endpoint and authentication to a Bedrock API key or AWS credentials. Most other code stays the same.

What APIs does GPT-5.5 support on Bedrock?
Only the Responses API at GA. Console support is coming soon, but other Bedrock APIs (Messages, Converse) are not supported for OpenAI models.

Is Codex on Bedrock the same as Kiro?
No. Codex is OpenAI’s coding agent. Kiro is AWS’s spec-driven coding platform running primarily on Claude. They are different products with different workflows. See our complete Kiro guide for the comparison.

Does Codex on Bedrock work in JetBrains and VS Code?
Yes. Codex on Bedrock has IDE integrations for VS Code, JetBrains, and Xcode, plus the Codex CLI and Codex App.

Can I use GPT-5.5 with AWS Bedrock AgentCore?
Yes. AgentCore is model-agnostic. You can build agents on AgentCore that use GPT-5.5, Claude Opus 4.8, or any other Bedrock model. See our AgentCore guide for details.

What regions support GPT-5.5 on Bedrock?
US East (N. Virginia), US West (Oregon), select EU regions at launch. GovCloud (US-West) has GPT-5.4 as of June 3, 2026. Additional regions are rolling out through 2026.

Is GPT-5.5 better than Claude Opus 4.8 for coding?
Mixed. Claude Opus 4.8 leads on SWE-bench Verified (88.6%) with strong honesty and code review improvements. GPT-5.5 leads on Terminal-Bench 2.1 (83.4% under Codex CLI). Test both on your specific codebase before committing.

Summary

The arrival of GPT-5.5, GPT-5.4, and Codex on Amazon Bedrock is the most consequential AI infrastructure change of 2026 for enterprise AWS users.

For the first time, you can use OpenAI’s frontier models inside your existing AWS governance with the same security, compliance, and procurement controls you already have for every other AWS service. Pricing matches OpenAI direct. Usage counts toward AWS commitments. No new vendor relationship to manage.

If you are on AWS and you have been blocked from using OpenAI because of compliance, security, or procurement friction, that block is gone. If you are running a multi-model strategy, you now have one API surface for both OpenAI and Anthropic models.

Test both GPT-5.5 and Claude Opus 4.8 on your specific workloads. Pick the right model for each task. Use Bedrock’s governance to control cost, audit access, and protect data. This is the playbook for enterprise AI in 2026.