OpenAI Models Are Now on AWS Bedrock — Here's Exactly What's Available and What's Coming
GPT-5.4 is live on Bedrock now. GPT-5.5 and Codex are coming within weeks. Here's the full availability timeline and what it means for Bedrock users.
GPT-5.4 Is Live on AWS Bedrock Right Now
GPT-5.4 entered limited preview on AWS Bedrock on April 28, 2025. GPT-5.5 is coming within weeks. Codex is available through AWS infrastructure. Amazon Bedrock’s managed agents platform has been rebranded as “powered by OpenAI.” If you’ve been building on Bedrock and routing around OpenAI because it wasn’t natively available, that workaround is now obsolete.
This isn’t a roadmap item or a partnership announcement with a vague timeline. The models are there. You can request access today.
The timing is not accidental. OpenAI and Microsoft restructured their partnership on April 27. Microsoft’s exclusive cloud arrangement was dissolved. OpenAI announced AWS availability the next day. One day. That’s how long it took for OpenAI to act on the new freedom.
AWS CEO Matt Garman explained the demand plainly: “This is what our customers have been asking for for a really long time. Their production applications run in AWS, their data is in AWS, they trust the security of AWS.” That’s not marketing language — that’s a description of the actual friction that existed before this week.
Why Bedrock Builders Wanted This
Coding agents automate the 5%. Remy runs the 95%.
The bottleneck was never typing the code. It was knowing what to build.
The practical problem was straightforward. If you’re an enterprise running production workloads on AWS, you had a choice: use the models your infrastructure already trusts, or use OpenAI models and accept a split architecture. Many teams defaulted to Anthropic’s Claude because it was already on Bedrock and the security posture was clean. OpenAI was available, but through workarounds that created real operational friction — and, as it turns out, threatened Microsoft with enough legal exposure that lawsuit language entered the conversation.
That’s the Signal quote worth sitting with: “Many companies defaulted to Anthropic/Claude because they were already on Bedrock — this is huge for OpenAI model accessibility.” This wasn’t about model quality. It was about where the data lives and who controls the security boundary.
Now, for the first time, GPT models run on Bedrock the same way they’ve run on Azure. No workarounds. No split architecture. The same managed infrastructure, the same IAM controls, the same VPC configurations your team already has in place.
If you’re comparing model performance for specific tasks — say, GPT-5.4 vs Claude Opus 4.6 on coding and document processing — you can now make that comparison inside a single cloud environment and deploy whichever wins without changing your infrastructure.
What’s Actually Available and When
Here’s the current state, as specifically as the announcements allow:
GPT-5.4 — Limited preview on Bedrock as of April 28. “Limited preview” means you may need to request access rather than flip a switch. AWS typically gates these rollouts by region and account type before general availability. If you’re on an enterprise agreement, that’s your fastest path.
GPT-5.5 — Announced as “coming within weeks” from the April 28 date. Given that GPT-5.5 is already the model people are shifting behavior toward — GPT-5.5 uses 72% fewer output tokens than Opus 4.7 on equivalent tasks, which has real cost implications at scale — this is the one to watch. Weeks means weeks, not quarters.
Codex — Available through AWS infrastructure now. Codex is OpenAI’s coding-focused model, and its availability on Bedrock matters specifically for teams building agentic coding workflows. The managed agents platform on Bedrock, now branded “powered by OpenAI,” uses OpenAI’s harnesses and models — making it functionally similar to the managed agents OpenAI introduced with their Frontier platform in February.
What’s not confirmed yet: Pricing on Bedrock for these models hasn’t been published in detail. AWS Bedrock typically applies its own pricing layer on top of model costs, and the token economics will matter. For context on what you’re working with, token-based pricing for AI models explains how to estimate costs before you commit a workload.
Getting Access: The Actual Steps
Step 1: Check your AWS region. Bedrock model availability is region-specific. GPT-5.4 limited preview will almost certainly start in us-east-1 and us-west-2 before expanding. Log into the Bedrock console and check the model catalog for your region. Now you know whether you’re in scope for the current rollout.
Step 2: Request model access. In the Bedrock console, navigate to Model Access. OpenAI models will appear as a provider once they’re live in your region. Request access — for limited preview models, this may require a use case description. Enterprise accounts with existing AWS support contracts can escalate through their TAM. Now you have an access request in flight.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Step 3: Validate your IAM permissions. Bedrock model invocation requires bedrock:InvokeModel permissions. If you’ve been using Claude on Bedrock, you likely already have this. Confirm the policy covers the new OpenAI model ARNs — these will follow the standard Bedrock ARN pattern (arn:aws:bedrock:region::foundation-model/openai.gpt-5-4 or similar). Now you have the right permissions scoped.
Step 4: Test with the Bedrock API. The Bedrock InvokeModel API is the same regardless of model provider. The request body format will follow OpenAI’s standard chat completions structure, wrapped in Bedrock’s envelope. If you’ve used Bedrock with any other model, the integration pattern is identical. Run a basic completion call. Now you have a working integration.
Step 5: Evaluate the managed agents platform. The “powered by OpenAI” rebrand of Bedrock’s managed agents platform is worth a separate evaluation pass. If you’re building multi-step agentic workflows, this gives you OpenAI’s harness running on Bedrock’s infrastructure. That’s a different value proposition than just calling the model directly. For teams already using MindStudio to chain models and agents visually — it supports 200+ models and 1,000+ integrations — this Bedrock availability means GPT models can now be part of AWS-native workflows without the architecture compromises that previously made that painful. Now you have a clear picture of which integration path fits your use case.
The Failure Modes to Expect
Limited preview access lag. “Limited preview” is doing real work in that announcement. Don’t assume your account has access until you’ve confirmed it in the console. Teams that plan a migration around this and hit an access wall will lose weeks. Request access now, before you need it.
Pricing surprises. AWS Bedrock adds a markup to model costs. GPT-5.5 is already priced at $5 per million tokens input and $30 per million tokens output through OpenAI’s API directly. Bedrock’s pricing layer will sit on top of that. Run the math before you migrate a high-volume workload. The cost structure that made sense on Azure may not translate directly.
Region availability gaps. If your data residency requirements pin you to a specific AWS region, verify that region has the model before designing around it. EU regions in particular tend to lag US availability on new models.
Managed agents platform maturity. The “powered by OpenAI” managed agents platform is new. Rebranding an existing platform and integrating a new model provider’s harnesses is non-trivial. Expect rough edges in the first weeks — logging, error handling, and retry behavior are the places where new integrations typically show gaps. Build in extra testing time if you’re using the managed agents layer rather than direct model invocation.
The Codex-specific consideration. Codex on AWS is interesting for agentic coding workflows, but the same harness detection issues that caused problems for Anthropic users recently apply here in a different form. If you’re building coding agents that reference specific tooling in their context, test carefully. The comparison of GPT-5.4 Mini vs Claude Haiku for sub-agent tasks is relevant if you’re thinking about where Codex fits in a multi-model agent stack.
The Bigger Picture You Should Care About
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Sam Altman described OpenAI’s current position as: “We have become an AI inference company now.” That’s a meaningful statement. It signals that OpenAI is no longer thinking of itself as a model lab that happens to have an API — it’s thinking of itself as infrastructure.
The Rezo framing from the source coverage captures the structural reality: “OpenAI has grown too big for any single cloud to fully serve.” That’s not a criticism of Microsoft. It’s a description of scale. When your inference demand exceeds what one cloud can serve, multi-cloud isn’t a strategy choice — it’s an operational requirement.
For you as a builder, this means the model selection conversation is now decoupled from the cloud selection conversation in a way it wasn’t before. You can choose OpenAI models and still run on AWS. You can run a mixed fleet — Claude for some tasks, GPT-5.5 for others — inside a single Bedrock environment. The comparison of how Anthropic, OpenAI, and Google are each betting differently on AI agents is worth reading in this context, because the infrastructure availability now matches the strategic divergence.
Microsoft’s position in all of this is worth understanding clearly. They retain a non-exclusive license to OpenAI’s IP and models through 2032. They remain a 27% shareholder. They locked in a 20% revenue share through 2030 — previously set to drop to 8%, so the new deal is actually more favorable to Microsoft than the old trajectory. The AGI clause that would have voided the deal if OpenAI declared AGI has been removed. Microsoft didn’t lose this negotiation. They traded exclusivity for financial certainty and removed the clause that gave OpenAI a unilateral exit.
What that means practically: Azure isn’t going away as an OpenAI deployment target. Models still release on Azure first — the exclusivity window for new model launches hasn’t been publicly specified, but it exists. If you need day-one access to new OpenAI models, Azure remains the faster path. If you need AWS-native deployment with existing Bedrock infrastructure, you now have a clean option that didn’t exist before.
What to Do This Week
The access request is the only time-sensitive action. Limited preview slots fill, and teams that request early get access earlier. Log into the Bedrock console, find the model catalog, and submit the access request for GPT-5.4 today.
While you’re waiting for access, do the cost modeling. Pull your current token usage from whatever model you’re running on Bedrock, apply the GPT-5.5 pricing (input and output separately — the output cost is where the difference shows up), and add an estimated Bedrock markup. If the numbers work, you have a migration case. If they don’t, you have a clear answer.
The managed agents platform deserves a separate evaluation, but not this week. Let it stabilize for a few weeks before building production workflows on it. Direct model invocation is the lower-risk starting point.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
One thing worth building toward: if you’re doing spec-driven development and generating full-stack applications from structured requirements, tools like Remy compile annotated markdown specs into complete TypeScript backends, databases, auth, and deployment — the kind of workflow where having GPT-5.5 available natively on your cloud infrastructure, rather than through a split architecture, actually changes what’s practical to build. Model availability on your existing infrastructure removes a category of friction that used to require architectural compromise.
The access request takes five minutes. The cost model takes an afternoon. Both are worth doing before GPT-5.5 hits general availability and the decision becomes urgent.
