Bedrock vs OpenAI for Enterprise in 2026: Which Models, Which Contract, Where the Data Stays

OpenAI's models have run on Amazon Bedrock since June 2026, so the platform decision is no longer about which models you can use. What it is about now (data handling, identity, lifecycle and lock-in) and a short plan for teams that built on OpenAI directly.

What changed

When we first wrote this piece, the honest summary was: OpenAI had models Bedrock didn't, and Bedrock had controls OpenAI's API didn't. On June 1, 2026 AWS made GPT-5.5, GPT-5.4 and Codex generally available on Bedrock. GPT-5.6 Sol, Terra and Luna followed in July, and in September the GPT-6 family: Astra, Sol, Luna and then GPT-6.1 Sol. The models are the same ones; what differs is the path to them.

The decision now

For a team on AWS, Bedrock is the default path to OpenAI models for the same reasons it always was for Claude: IAM instead of another API key, VPC endpoints instead of the public internet, CloudTrail for every call, the spend on the bill you already reconcile, and evidence your auditors already accept. The reasons to go direct are narrower than they were:

• You need an OpenAI model or feature before it reaches Bedrock. AWS describes Bedrock Managed Agents powered by OpenAI as coming soon, so some capabilities do arrive later.

• You depend on OpenAI's own hosted tools or product surface rather than the models.

• You aren't on AWS, in which case Bedrock's advantages mostly don't apply.

Three details that bite

API support is per model

OpenAI models arrived on Bedrock through the Responses API. In August 2026 AWS added Converse and Chat Completions on the bedrock-runtime endpoint for the GPT-5.6 models. Code written against the OpenAI SDK generally moves with a base URL and credential change where Chat Completions or Responses are supported; code on Converse gets the benefit of swapping between Claude, Nova and GPT with one parameter. Check the model card for the exact model you use.

Whose deprecation schedule applies

Most Bedrock models carry Bedrock's own lifecycle dates. GPT-6.1 Sol's model card says it follows OpenAI's first-party deprecation terms instead, with no separate Bedrock minimum-availability date. Record the policy for each model when you adopt it — we wrote up how Bedrock retirements catch teams out in our post on Claude Sonnet 4's end of life.

Cross-Region inference and residency

Geo and Global cross-Region inference arrived for OpenAI models in August. Global profiles give the most capacity and route requests to any commercial Region where the model runs; Geo profiles keep routing within a geography such as the US. If your data must stay in one country, choose the profile deliberately — the default in an SDK example is often global.

If you built directly on OpenAI

1. Put every model call behind one internal client, if it is not already. That is the change that makes every later step a configuration change.

2. Build a small evaluation set from your own traffic and run it against the same model on Bedrock — results should match, and this is where any API-shape difference shows up.

3. Move one workload behind a flag, compare latency, errors and cost for a week, then move the rest.

4. Once you are on Bedrock, run the same evaluation set against Claude and a smaller model for each task. Most teams find at least one workload that gets cheaper or better.