OpenAI finally brings GPT, Codex, and Managed Agents to AWS — yes, really

OpenAI just did something that feels both overdue and genuinely useful: they put GPT models, Codex, and Managed Agents directly into AWS. Not through some janky third-party integration or a half-baked API wrapper. Natively, in the AWS ecosystem.

If you’ve been building AI applications on AWS, you know the drill. You either hit OpenAI’s API directly and deal with data leaving your VPC, or you go with something like Amazon Bedrock and get a narrower set of models. This announcement changes that calculus.

What’s actually available

The three big pieces are GPT models (the full chat and completion family), Codex (the code generation model that powers GitHub Copilot), and what OpenAI calls Managed Agents. The agents piece is interesting — it’s essentially a hosted version of the agent framework they’ve been dogfooding internally. You define tools, give it a goal, and it runs in your AWS account with IAM roles controlling what it can touch.

All of this runs inside your VPC. That’s the headline. Data doesn’t leave AWS. If your compliance team has been blocking OpenAI because of data residency or privacy concerns, that argument just got a lot weaker.

The pricing question

OpenAI hasn’t published exact pricing for the AWS deployment yet, which is annoying. They say it’s “competitive with direct API access” but that usually means a premium for the convenience of staying inside AWS. My guess is 10-20% markup over the standard API rates, which is honestly reasonable for not having to build your own infrastructure layer.

Why this matters more than it sounds

Amazon Bedrock already has models from Anthropic, Meta, Cohere, and a bunch of others. But GPT-4o and Codex are the ones most enterprises actually want. The fact that they’re now available as a managed service inside AWS means you can skip the whole “should we build a proxy layer” conversation.

I’ve seen too many teams spend months building custom routing, caching, and monitoring for OpenAI’s API when they could have just used a managed service. This eliminates that busywork.

Managed Agents are the dark horse

The agents part is what I’m most curious about. OpenAI has been quietly building out their agent infrastructure for a while, and putting it inside AWS with native IAM integration is a smart move. You can give an agent read access to an S3 bucket, write access to a DynamoDB table, and let it figure out the rest.

Will it hallucinate and delete your production database? Probably not if you scope the IAM policy correctly. But I’d still start with read-only permissions and work up.

What’s missing

No fine-tuning support at launch. If you wanted to customize GPT-4o on your own data, you still need to use OpenAI’s platform directly. Also no batch inference pricing yet. And the agents are limited to 10 concurrent runs per account unless you negotiate a higher limit.

These are real limitations, but for the use case of “I need GPT inside my AWS account with data staying put,” this covers it.

The bottom line

This isn’t revolutionary. It’s a distribution deal. But it’s a distribution deal that removes the biggest blocker for enterprise adoption: data leaving the cloud. If you’re already in AWS and you’ve been eyeing OpenAI’s models but held back on security grounds, your excuse just evaporated.

Go build something. Just watch those IAM policies.