Azure OpenAI vs OpenAI API vs AWS Bedrock

For most of the last three years, this comparison had a simple shape: OpenAI models lived on the OpenAI API and Azure, and Bedrock was where you went for Claude and open models. That is no longer true. Amazon made OpenAI’s GPT-5.5, GPT-5.4, and Codex generally available on Bedrock in June 2026, with pricing that matches OpenAI’s direct rates and usage that counts toward existing AWS commitments [3]. When the GPT-5.6 family launched on July 9, 2026 [15], Microsoft made it generally available in Foundry the same day [1], and AWS followed on Bedrock on July 13 [2].

So the question has changed. It is less “which platform has the model” and more “which operating envelope fits my product”: where data is processed, who can see it, which identity and network controls apply, which features and limits you actually get, and whose invoice it lands on.

The short answer: build on the cloud where your data, identity, and spending commitments already live, unless you need OpenAI’s newest features and limits the day they ship, in which case go direct. Either way, put a thin gateway between your product and the provider so the choice stays reversible.

This guide is part of our series on adding AI features to an existing product.

Side-by-side comparison

The rest of this guide explains the rows that most often decide the outcome.

OpenAI API: the reference implementation

Going direct is the simplest integration and gives you OpenAI’s full feature set and limits first. Every other platform’s OpenAI offering is, in effect, measured against it.

Data handling. OpenAI states that it does not train on API data by default [7]. API inputs and outputs may be retained for up to 30 days for abuse monitoring, and zero data retention is available for qualifying use cases on eligible endpoints [8]. Enterprise customers approved for advanced data controls can create projects pinned to a region, where requests are handled in-region and model requests and responses are not stored at rest [7].

Where it wins. Teams that want new models, features, context limits, and service tiers the moment they ship. Smaller companies without cloud-bound procurement rules. Products where OpenAI is the only model provider you plan to use.

Tradeoffs. It is another vendor for your security team to review, another invoice outside your cloud commitments, and another identity system. Enterprise customers who have standardized on Azure or AWS will often ask why their data leaves that cloud, and “because the API was easier” is not a satisfying answer in a security review.

Azure OpenAI in Microsoft Foundry: the Microsoft-native path

Azure OpenAI now sits inside Microsoft Foundry, alongside models from other providers. For companies already running on Azure, it keeps AI traffic inside the same tenant, identity model, and billing relationship as everything else.

Data handling. Microsoft’s documentation is explicit: prompts, completions, embeddings, and training data are not available to other customers or to OpenAI, are not used by model providers to improve their models, and are not used to train foundation models without your permission. Microsoft hosts the models in Azure, and they do not interact with services OpenAI operates [5].

Two details deserve attention in a security review. First, deployment type controls where processing happens: Global deployments may process prompts in any geography where the model is deployed, while Data Zone deployments keep processing within the zone, such as anywhere in the United States or within EU member states. Data at rest stays in your designated geography either way [5]. Microsoft added an Asia-Pacific Data Zone in July 2026 [1]. Second, abuse monitoring can select samples of prompts and completions for review, including human review when content is flagged; customers approved for modified abuse monitoring turn off that storage and human review, though automated checks can still run [5]. Guardrails, previously called content filters, evaluate prompts and completions in real time against configured thresholds [5], which occasionally means tuning them for legitimate domain content.

Where it wins. Microsoft-centric companies with Entra ID, Azure networking, Microsoft 365 integrations, and committed Azure spend. Microsoft made GPT-5.6 available on day one across Global Standard and Global Priority Processing in its existing global regions, Data Zones Standard, and Global Provisioned deployments [1], so the historical lag behind OpenAI has narrowed sharply. Provisioned throughput gives predictable capacity for steady, high-volume workloads.

Pricing notes. Standard global pricing generally tracks OpenAI’s rates, but not always. Microsoft announced promotional GPT-5.6 Sol pricing of $4 per million input tokens and $20 per million output tokens from September 1 through at least November 30, 2026, below its standard $5 and $30 rates, and directs customers to sales for Data Zone and Priority Processing prices [1]. Check the current page before modeling costs; see LLM API costs in production for how to build that model.

Tradeoffs. Quota is managed per subscription and region, so capacity planning becomes part of the job. Not every model or feature reaches every region or deployment type at once. And Azure’s portal and documentation have been through several renames, which makes it harder for new team members to find the right page.

Amazon Bedrock: the multi-vendor AWS path

Bedrock has become the broadest single catalog of the three: OpenAI’s frontier models now sit next to Claude, Amazon’s own models, and a long list of open-weight models from providers such as DeepSeek, Qwen, and Moonshot [16].

Data handling. AWS states that Bedrock does not store or log prompts and completions, does not use them to train AWS models, and does not distribute them to third parties. Each model provider’s model runs in a deployment account operated by the Bedrock team, and providers have no access to those accounts, logs, or customer prompts [6].

Where it wins. AWS-centric companies whose data already sits in S3, RDS, or other AWS services, that want IAM, KMS, and PrivateLink applied to AI traffic, and that want OpenAI and Claude available in one place with spend counting toward AWS commitments [3]. Bedrock’s inference options are clear for residency planning: In-Region keeps requests in a single region, Geo cross-region routes within a geography such as the US or EU, and Global routes anywhere for maximum throughput [10]. Its service tiers cover the full range, from Flex for work that can wait, through Standard and Priority, to Reserved capacity for workloads that cannot tolerate interruption [9].

Public sector. In June 2026, AWS announced FedRAMP High and DoD IL-4/5 approval for OpenAI GPT models on Bedrock in AWS GovCloud (US) [12]. Azure has offered OpenAI models to government customers through Azure Government for longer, with FedRAMP High authorization dating to 2024 [13]. If you sell to federal agencies, the right platform is usually whichever one already sits inside your authorization boundary.

Tradeoffs. OpenAI on Bedrock is new, and parity is not guaranteed. An independent test of GPT-5.6 on Bedrock shortly after launch found a documented context window of 272K tokens compared with 1.05M for first-party OpenAI access, and several service tier and cross-region options unavailable in the tested configuration [4]. Treat that as a dated snapshot that AWS may have changed since, but also as a reason to verify limits for your exact model, region, and tier before committing.

Watch who actually operates inference

The same model name can mean different operators, data paths, and contracts depending on how you buy it. Claude is a useful example. Anthropic distinguishes between partner-operated platforms, where the cloud provider runs inference and invoices you, such as Amazon Bedrock and Google Cloud, and Anthropic-operated offerings billed through a cloud marketplace, such as Claude Platform on AWS and Claude in Microsoft Foundry [11]. Pricing can differ too: Claude’s regional endpoints on Bedrock and Google Cloud carry a 10% premium over global endpoints, while US-only inference on Anthropic’s own API applies a 1.1x multiplier [11].

For your security review and your customers’ questionnaires, record three things for every model you use: who operates inference, which company’s data processing terms apply, and where abuse monitoring data, if any, is stored.

Do not assume feature parity

Even when a model is available everywhere, the surrounding details often differ: maximum context, output limits, caching behavior, service tiers, batch support, structured output and tool calling behavior, and regional availability. The practical defense is the same discipline you use for model upgrades:

• Keep a capability matrix per provider covering the limits and features your product actually uses.
• Run your eval suite against each provider and region you deploy to, not just against the model name.
• Enforce token budgets per provider, so a prompt that fits one platform’s context window does not fail on another.
• Re-run the suite when a provider announces changes, since limits and tiers on all three platforms moved repeatedly in 2026.

Build for portability

The good news is that switching costs are falling. Bedrock exposes OpenAI models through the OpenAI Responses API on its Mantle endpoints [14], and Azure supports the Responses API in Foundry as well [5]. In many cases, moving between platforms starts with changing a base URL and credentials.

Starting, not finishing. Authentication, quotas, error formats, content filtering, and the limits above still differ. That is why we recommend a thin internal gateway between your product and any provider: one interface for model calls, provider-specific configuration for limits and credentials, and logging of model, provider, tokens, latency, and cost on every request. A gateway also makes a second provider practical as a fallback during outages, which we cover in making LLM features reliable.

How to choose

If none of these clearly applies, pick based on where your enterprise customers expect their data to go. That expectation will surface in every security review you do.

Questions your customers’ security teams will ask

Have written answers ready for each provider you use:

• Which company operates inference, and which data processing agreement applies?
• Is customer data used for training, and is that commitment contractual?
• Where are prompts and completions processed, and where is any data stored at rest?
• Is any content retained for abuse monitoring, for how long, and can a human review it?
• Which identity, encryption, and private networking controls protect AI traffic?
• Which compliance attestations cover the specific model and region in use?
• Can administrators disable AI features for their organization?

Our enterprise RAG security and governance guide and enterprise AI governance guide cover how to turn these answers into controls.

Keep reading

• How to add AI features to an existing software product (the pillar for this series)
• LLM API costs in production: how to estimate and cut them
• Choosing an LLM for your product
• Latency, rate limits, and fallbacks: making LLM features reliable

CMW AI Labs builds production AI features on Azure, AWS, and direct model APIs, with gateways that keep the platform decision reversible. If you are weighing where your first AI feature should run, talk to our team.

Sources

1. Microsoft Azure Blog, “GPT-5.6 now available in Microsoft Foundry,” July 9, 2026 (updated August 24, 2026). https://azure.microsoft.com/en-us/blog/gpt-5-6-now-available-in-microsoft-foundry/
2. AWS What’s New, “OpenAI GPT-5.6 Sol, Terra, and Luna now generally available on Amazon Bedrock,” July 13, 2026. https://aws.amazon.com/about-aws/whats-new/2026/07/openai-gpt-sol-terra/
3. IT Brief Australia, “Amazon adds OpenAI models to Bedrock for AWS users,” June 2, 2026. https://itbrief.com.au/story/amazon-adds-openai-models-to-bedrock-for-aws-users
4. Let’s Data Science, “OpenAI GPT-5.6 Models Reach General Availability on Amazon Bedrock” (summarizing independent testing by Classmethod), July 15, 2026. https://letsdatascience.com/news/openai-gpt-56-models-reach-general-availability-on-amazon-be-b2fa80b8
5. Microsoft Learn, “Data, privacy, and security for Models sold by Azure in Microsoft Foundry.” https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacy
6. AWS, “Data protection,” Amazon Bedrock User Guide. https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html
7. OpenAI, “Expanding data residency access to business customers worldwide.” https://openai.com/index/expanding-data-residency-access-to-business-customers-worldwide
8. OpenAI, “How we use your data,” API documentation. https://platform.openai.com/docs/models/how-we-use-your-data
9. AWS, “Service tiers for optimizing performance and cost,” Amazon Bedrock User Guide. https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html
10. AWS, Amazon Bedrock model cards, “Regional availability” section (In-Region, Geo, and Global inference). https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-premier.html
11. Anthropic, “Pricing,” Claude Platform Docs (accessed September 2026). https://platform.claude.com/docs/en/about-claude/pricing
12. AWS What’s New, “OpenAI GPT, OpenAI GPT OSS, and NVIDIA Nemotron models on Amazon Bedrock receive FedRAMP High and DoD IL-4/5 approval in AWS GovCloud (US),” June 25, 2026. https://aws.amazon.com/about-aws/whats-new/2026/06/addl-bedrock-model-fedramp-il-5-govcloud
13. Microsoft, “Azure OpenAI, including GPT-4o, Approved as a Service within the FedRAMP High Authorization for Azure Government,” Azure Government blog, August 12, 2024. https://devblogs.microsoft.com/azuregov/azure-openai-fedramp-high-for-government/
14. AI Engineer Guide, “How to use OpenAI models on AWS Bedrock.” https://aiengineerguide.com/til/openai-models-on-aws-bedrock/
15. Nils Durner, “GPT 5.6 release,” July 2026. https://ndurner.substack.com/p/gpt-56-release
16. AWS What’s New, “Amazon Bedrock adds support for six fully-managed open weights models,” February 10, 2026. https://aws.amazon.com/about-aws/whats-new/2026/02/amazon-bedrock-adds-support-six-open-weights-models
17. Let’s Data Science, “OpenAI Cuts GPT-5.6 Luna and Terra Prices,” July 2026. https://letsdatascience.com/news/openai-cuts-gpt-56-luna-and-terra-prices-ae12009d