GPT-5.5 just landed on AWS, and it’s open to everyone not just enterprise accounts, not just preview waitlists. All users. That’s the headline. But if you’re trying to figure out what this actually means for your stack, your costs, or your current OpenAI API setup, here’s the real breakdown.Now I have everything I need. Here’s the full article:
GPT-5.5 is live on Amazon Bedrock for all users no waitlist, no preview request, no special enterprise agreement required. As of June 1, 2026, you can run GPT-5.5 in production workloads on Amazon Bedrock with the same security, governance, and operational controls you already use across AWS. That’s a bigger deal than the headline sounds. Let me break down why and what you actually need to do to use it.
What Is GPT-5.5 on AWS And Why Does It Matter?
On April 28, 2026, AWS and OpenAI confirmed a $50 billion partnership that effectively ends OpenAI’s exclusivity with Microsoft Azure. That partnership is what made this possible. For years, if you wanted GPT-class models and you were an AWS shop, you had a problem: your AI layer lived in Azure or went directly through OpenAI’s API, completely separate from your IAM policies, your CloudTrail logs, your VPC setup, your AWS cost commitments. That forced a lot of teams into awkward dual-cloud setups they didn’t really want.
Now, customers can access GPT-5.5 the most advanced frontier model from OpenAI along with GPT-5.4 and Codex on Amazon Bedrock, with pricing that matches OpenAI first-party rates and no additional fees, and usage counting toward existing AWS commitments.
So the short version: same model, same pricing, but now it lives inside your AWS account. No extra vendor. No separate billing. No new security perimeter to manage.
Honestly, I didn’t expect AWS and OpenAI to pull this off this fast. The partnership was announced in April, and GA hit June 1st. That’s a quick turnaround for something this operationally significant.
GPT-5.5 vs GPT-5.4 on Bedrock Which One Do You Actually Need?
Here’s where most people get confused. There are two OpenAI models on Bedrock now, and picking wrong costs you either money or performance.
GPT-5.5 scored 82.7% on Terminal-Bench 2.0 and is OpenAI’s “strongest agentic coding model to date.” It’s the one you reach for when the task genuinely needs maximum capability: complex multi-step reasoning, hard agentic coding workflows, professional analysis pipelines. Think law firms doing long-form document analysis, quant teams running multi-variable modeling, or engineering teams building autonomous code agents.
GPT-5.4 is what AWS calls the “best price-performance” option. It handles most coding, reasoning, and professional tasks extremely well just at lower cost. If you’re building internal tools, running moderate agentic workflows, or doing most standard enterprise workloads, GPT-5.4 is probably the smarter pick.
The honest truth? Most teams don’t need GPT-5.5 for everything. I’ve seen this pattern before with Claude Opus versus Sonnet. People default to the top model and then wonder why their costs exploded. Start with GPT-5.4, test your actual workload, and only move up to GPT-5.5 for the tasks that truly need it.
GPT-5.5 Now Available on AWS for All Users The Exact Setup Steps
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. This is the detail that trips most teams up in the first hour. If you have existing Bedrock code using the Converse API or the Messages format, you can’t just swap in the model ID and expect it to work. You need to migrate to the Responses API endpoint.
Here’s the practical walkthrough:
Step 1 Request model access
Sign into the AWS Console, go to Amazon Bedrock, and navigate to Model Access. Find the OpenAI GPT-5.5 entry and submit an access request. Approval is typically instant for most accounts, but be aware that model access is regional. If you need GPT-5.5 in a specific region, you have to request access in that region specifically.
Step 2 Check your region
GPT-5.5 is currently available in the US East (Ohio) region. GPT-5.4 is available in US East (Ohio) and US West (Oregon). If you tried calling GPT-5.5 from Oregon and got a 404, that’s why. As of June 2, 2026, GPT-5.5 is only available in Ohio. EU and other regions will lag, so if data residency matters for your compliance situation, verify before you architect anything around it.
Step 3 — Set up your environment
Configure your environment with your Bedrock API key and set the base URL to the regional Bedrock mantle endpoint:
OPENAI_BASE_URL=”https://bedrock-mantle.us-east-2.api.aws/openai/v1″
The model ID you pass is openai.gpt-5.5 — not just gpt-5.5. That prefix matters.
Step 4 — Use the Responses API
You’re calling client.responses.create(), not client.chat.completions.create(). If you’re coming from existing OpenAI API code, that’s the migration point. The OpenAI Python SDK still works you just point base_url at the Bedrock endpoint and use your Bedrock API key as the api_key.
Step 5 Validate and run
Start with a simple test call before you move any production workload over. Verify logging is going to CloudTrail, confirm your IAM permissions are scoped correctly, and check your budget alerts are set. Don’t skip this step — token costs on GPT-5.5 add up faster than you expect if something misbehaves.
What About Codex on Bedrock?
More than 4 million developers use Codex every week to write, refactor, debug, test, and validate code across large codebases. With GPT-5.5 powering inference, Codex introduces a new class of intelligence optimized for complex, long-horizon developer workflows.
Codex is available through the Codex App, CLI, and IDE integrations, with all inference routed through Amazon Bedrock and the same security protections including IAM, VPC isolation, and encryption that customers already use on AWS.
So you can configure Codex to run inference through Bedrock instead of directly through OpenAI. For enterprise engineering teams where data governance is non-negotiable, this matters. Code goes through your AWS account, not out to OpenAI’s endpoints directly.
The IDE integrations cover Visual Studio Code, JetBrains, and Xcode. If your team is already using Codex, switching the inference backend to Bedrock is mostly a configuration change, not a workflow change. That’s the cleanest upgrade path.
The Real Benefits Nobody’s Spelling Out
People are writing about this like it’s just “OpenAI models on a different cloud.” That’s missing what actually changed.
AWS commitment spend now covers AI costs. If your organization has AWS Enterprise Discount Program commitments or Reserved Instance deals, your GPT-5.5 usage counts toward those. That can meaningfully reduce effective cost if you’re already spending heavily on AWS. Finance teams love this because it simplifies budget planning and eliminates a separate OpenAI vendor relationship.
IAM-native access control. You can scope exactly which roles, teams, and services can call GPT-5.5, using the same IAM policies you use for everything else. No more managing separate OpenAI org-level API keys and hoping they’re rotated properly. This is huge for security teams who’ve been dealing with shadow AI risks in enterprise environments.
CloudTrail logging for every inference call. Every GPT-5.5 call through Bedrock shows up in your audit logs. That’s critical for regulated industries finance, healthcare, legal where you need to demonstrate that you know exactly what data went to which AI system and when. If you’ve been building out an AI incident governance playbook, this is the kind of observability you’ve been waiting for.
VPC isolation. You can route all traffic through PrivateLink and never let GPT-5.5 calls touch the public internet. For teams with strict data handling requirements, this is the difference between “we can’t use this” and “we can actually deploy this.”
The Catch (Because There’s Always a Catch)
Look, I’ve spent enough time on Bedrock to know the gaps don’t always make it into the announcement posts.
Context window is capped at 272K tokens on Bedrock. GPT-5.4 and GPT-5.5 have a 272,000-token context window on Amazon Bedrock, and Bedrock rejects requests that exceed this limit. The direct OpenAI API gives you 1M context on GPT-5.5. That’s a meaningful difference if you’re running long-form analysis on large documents or working with big codebases. Plan your chunking strategy before you hit this wall in production.
No console playground yet. 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. If you’re used to testing prompts in the Bedrock console before coding them up, you’ll be writing test scripts for now.
Hosted tools aren’t available. Hosted tools run through OpenAI-operated service infrastructure and are unavailable in the initial Amazon Bedrock offering. Client-side tool use works fine, but if your architecture depended on OpenAI’s hosted web browsing or code interpreter tools, those don’t transfer over to the Bedrock deployment.
Output tokens cost more than Claude. Output token pricing runs higher than Claude Opus 4.8 $30 per million output tokens versus $25 for Claude. For output-heavy workloads like long document generation or verbose reasoning traces, that difference compounds fast. Do the math on your actual token split before committing.
Managed Agents are still in preview. The companion product for fully managed agentic workflows on Bedrock is not yet generally available. If that’s a key part of your roadmap, it’s a future capability, not something you can ship against today.
Who Should Switch Right Now vs Wait
Real talk: not everyone needs to move immediately.
Switch now if: You’re an AWS-native team that’s been using OpenAI directly through the API and dealing with the split governance headache. The consolidation benefit is immediate. Also switch if you’re in a regulated industry where the CloudTrail logging and IAM controls unlock use cases you couldn’t touch before.
Wait if: You need EU data residency. Regional availability is limited right now and expanding. Building your production architecture around Ohio-only availability and then needing to migrate when EU regions open is a real pain you can avoid by waiting 60-90 days.
Test it first if: You have output-heavy workloads. Run your actual production token profile against the pricing before making any commitment. The math might still favor direct OpenAI for some use cases, particularly if you’re already on a custom API pricing agreement.
The Governance Angle (This Is Where It Gets Interesting)
The combination of GPT-5.5 plus Bedrock’s security primitives creates an opportunity a lot of organizations haven’t had before: genuinely enterprise-grade AI governance without architectural compromises.
Think about what this enables. You can set SCPs at the AWS Organizations level that prevent GPT-5.5 from being called without approved logging configurations. You can use VPC endpoint policies to lock down exactly which workloads can reach the model. You can feed CloudTrail events into your SIEM and get alerts when unusual inference patterns show up which is exactly the kind of silent behavioral drift detection that mature AI programs need.
If your organization has been working through AI risk classification frameworks, the Bedrock deployment path makes it dramatically easier to apply different access controls to different risk tiers. High-risk use cases get PrivateLink and KMS. Standard workloads get normal Bedrock endpoints. You manage it all from one control plane.
That’s a different conversation than “which model is best.” It’s about whether your AI deployment can actually meet the governance requirements your legal and compliance teams are asking for. With GPT-5.5 on Bedrock, for many teams, the answer just changed from “not yet” to “yes.”
GPT-5.5 on Bedrock vs Direct OpenAI API The Actual Decision
Here’s the framework I’d use:
Use Bedrock if:
- Your organization’s AWS commitments benefit from spend consolidation
- Compliance requires audit logging, VPC isolation, or IAM-native access control
- Your team manages AI governance centrally and needs one control plane
- You’re already running other models on Bedrock and don’t want another vendor
Stick with direct OpenAI API if:
- You need the full 1M context window (Bedrock caps at 272K)
- You rely on OpenAI’s hosted tools (web browsing, code interpreter)
- You have a custom pricing agreement with OpenAI that Bedrock’s standard rates don’t match
- You need EU or APAC regional availability that Bedrock doesn’t cover yet
There’s no universal right answer. The wrong move is defaulting to Bedrock just because it sounds more enterprise-friendly, or avoiding it because you haven’t looked at it yet. The right move is actually running your token volumes through both pricing paths and checking your compliance requirements against what each path provides.
What About AI Identity and Security on These New Endpoints?
One thing that doesn’t come up enough in these GA announcements: when you add a new frontier model endpoint to your stack, you’re also adding a new attack surface. Prompt injection, credential exposure through model outputs, and AI agent identity risks become more salient when you’re routing sensitive enterprise workloads through agentic systems.
Bedrock’s IAM controls help significantly with the authentication side. But the application-layer risks what your prompts contain, what your outputs log, how your agent decisions are audited those are still on you to get right. Worth having that conversation with your security team before you ship production workloads, not after.
And if you’re using Codex on Bedrock for autonomous code generation across large repositories, set human-approval checkpoints for anything that touches production infrastructure. The AI bias and governance controls conversation applies to code generation just as much as it does to decision-making systems.
The Fastest Way to Get Started Today
Don’t overthink it. Here’s the minimum viable path:
- Log into AWS Console → Amazon Bedrock → Model Access → request openai.gpt-5.5 in us-east-2
- Generate a Bedrock API key from the Bedrock console
- Install the OpenAI Python SDK (pip install openai)
- Set OPENAI_BASE_URL to https://bedrock-mantle.us-east-2.api.aws/openai/v1 and OPENAI_API_KEY to your Bedrock key
- Call client.responses.create(model=”openai.gpt-5.5″, …) and verify it returns correctly
- Set up a CloudWatch budget alert before you run anything at scale
The whole setup takes about 20-30 minutes if you have AWS console access already. The model access approval is instant for most accounts.
What you do after that depends entirely on your workload. But the access barrier is gone GPT-5.5, GPT-5.4, and Codex are now generally available on Amazon Bedrock, ready to deploy in production applications and agents today.
Start with a non-critical workload, validate your cost profile and governance setup, then expand from there. That’s the move that won’t bite you later.
