AWS made GPT-5.5, GPT-5.4, and Codex generally available on Amazon Bedrock, giving enterprises a governed path to OpenAI models through existing AWS secur
What Went GA on Bedrock
Amazon Bedrock now offers general availability for OpenAI's GPT-5.5 and GPT-5.4 models alongside Codex. General availability matters more than a preview badge: it signals that the models are covered by standard service terms, quotas, and support, which is the threshold most enterprises require before putting a model behind a customer-facing feature or an internal workflow that people depend on.
Having a general-purpose pair like GPT-5.5 and GPT-5.4 plus a code-focused option in Codex lets teams route work by task rather than forcing one model to do everything. Broad reasoning and drafting can go to the general models, while code generation, review, and refactoring can go to Codex.
Why "Governed Path" Is the Point
The real story is not that these models exist but that they are reachable through Amazon Bedrock. Instead of provisioning a separate vendor relationship, handling another set of API keys, and negotiating a new data-handling agreement, teams call OpenAI models through the same Bedrock interface they already use for other providers. Your requests stay inside your AWS account boundary rather than routing through an external endpoint you have to separately vet.
That means the controls you already run apply without new plumbing. Governance stops being a per-vendor exercise and becomes a property of the platform.
- Identity and access control through IAM roles and policies you already maintain.
- Request and response logging for audit and later review.
- Network isolation so traffic stays within your account boundary.
- Encryption and key management consistent with the rest of your AWS footprint.
- Centralized cost tracking on the same bill as your other AWS usage.
Practical Migration Guidance
If you are already calling OpenAI models directly, moving to Bedrock is mostly a matter of swapping the client and adjusting request formatting to the Bedrock invocation shape. Keep your prompts and evaluation set the same across the switch so you can confirm that behavior matches before you cut traffic over. Run both paths in parallel for a period, compare outputs on real inputs, and only retire the direct integration once the Bedrock path holds up.
Choosing between the models is worth deliberate testing rather than defaulting to the newest label. GPT-5.5 and GPT-5.4 give you room to trade capability against cost and latency for a given task, and Codex is the natural first choice for engineering workflows. Build a small benchmark from your own representative tasks and let the results, not the version number, drive the routing decision.
Who Benefits Most
The organizations that gain the most are the ones for whom procurement and compliance, not model quality, were the blocker. Regulated industries, security-conscious platform teams, and anyone standardized on AWS can now adopt OpenAI models without standing up a parallel governance regime. The approval conversation shifts from "can we trust a new external vendor" to "which Bedrock model fits this workload," which is a far shorter path to production.
For teams already committed to Bedrock as a model gateway, this widens the menu without widening the attack surface or the operational overhead. That combination of more choice and unchanged controls is what makes GA here worth acting on rather than just noting.