In this guide · 8 sections
- 01Quick comparison: which platform suits which business
- 02Data use and training: what each vendor commits to
- 03Abuse monitoring and data retention compared
- 04Regions near Singapore and data residency
- 05Model availability, billing and credits
- 06How to choose and test in an afternoon
- 07Do it yourself or get help?
- 08Frequently asked questions
Short answer
Use the OpenAI API for the newest OpenAI models and the simplest set-up. Use Azure if you already run on Microsoft and want Azure contracts, regions and billing. Use Amazon Bedrock if you are on AWS, want Claude and other models behind one API, and want Singapore-region endpoints.
Quick comparison: which platform suits which business
All three are serious, enterprise-grade options. The right one usually depends on where your other systems already live and which models you need.
OpenAI API
- Direct access to OpenAI's own models and features
- Simplest sign-up: an API key and prepaid or invoiced billing
- Abuse monitoring logs kept up to 30 days by default
- Singapore data residency for storage, by approval
Best for: Small teams that want OpenAI models with the least set-up.
Azure (Microsoft Foundry)
- OpenAI and other models under Microsoft's data protection terms
- Billed through your Azure subscription
- Choice of Global, Data Zone or regional deployment types
- More set-up: resource, region, deployment and quotas
Best for: Businesses already on Microsoft 365 or Azure.
Amazon Bedrock
- Claude, Amazon Nova and many other models through one API
- Asia Pacific (Singapore) endpoint, ap-southeast-1
- Configurable data retention modes, including zero retention
- Billed through your AWS account
Best for: Businesses on AWS, or that want Claude and model choice.
For token prices across providers, see our LLM API pricing comparison. This guide focuses on data, regions and set-up.
Data use and training: what each vendor commits to
The first question most SMEs ask is whether customer data sent to the model will be used to train it. All three say no by default for API use.
| Platform | Training on your API data | Source |
|---|---|---|
| OpenAI API | Not by default; you own inputs and outputs where allowed by law | OpenAI enterprise privacy page |
| Azure (Models sold by Azure) | Prompts, outputs and training data NOT available to OpenAI or other model providers, and NOT used to train foundation models without your permission | Microsoft Foundry data privacy page |
| Amazon Bedrock | Content not used to improve base models and not shared with model providers | AWS Bedrock FAQs and security page |
See OpenAI's enterprise privacy page, Microsoft's data, privacy and security page for Foundry models and the Amazon Bedrock FAQs. Training is only part of the picture: all three may still retain data for a period to detect abuse.
Abuse monitoring and data retention compared
Each vendor monitors for misuse, and that usually means keeping prompts and outputs for a limited time.
- OpenAI API: OpenAI's data controls page says abuse monitoring logs are generated for all API usage and kept for up to 30 days by default. Eligible customers can apply for Zero Data Retention or Modified Abuse Monitoring, subject to OpenAI's approval.
- Azure: Microsoft says prompts and completions are evaluated for harmful content and may be reviewed for code of conduct violations; managed customers can apply to modify abuse monitoring. For Global and Data Zone deployments, stored data, including the abuse monitoring store, stays in the customer-designated geography.
- Amazon Bedrock: Bedrock's data retention settings offer modes from zero retention (none) upwards. Some newer models require an aws_review mode, where prompts and outputs are kept within AWS for up to 30 days for human review by AWS, not the model provider.
Regions near Singapore and data residency
Where data is stored and where inference runs are separate questions, and the answers differ by platform.
| Platform | Singapore option | What it covers |
|---|---|---|
| OpenAI API | Singapore data residency (sg.api.openai.com) | Storage yes, processing no; needs approval for abuse monitoring controls; 10% uplift on models released from 5 March 2026 |
| Azure | Southeast Asia region for Foundry resources | Regional deployments process in the chosen geography; Global deployments may process wherever the model is deployed |
| Amazon Bedrock | Asia Pacific (Singapore), ap-southeast-1 | Content stored at rest in the Region you use; cross-Region profiles route within APAC or globally |
For Azure, Microsoft explains that for any deployment labelled Global, prompts and responses may be processed in any geography where the model is deployed, while Data Zone keeps processing within a defined zone. For Bedrock, AWS describes geographic cross-Region inference as keeping requests within boundaries such as APAC, and global profiles as routing worldwide. If your contracts or customers require processing in Singapore, check the exact model, region and deployment type before you build. Our guide to AI data residency explains why this matters under the PDPA.
Model availability, billing and credits
Models. OpenAI's API offers OpenAI's own models. Azure offers OpenAI models alongside models from other providers in Microsoft Foundry. Bedrock offers Anthropic's Claude, Amazon's Nova and models from other providers. Availability of a specific model version differs by platform, region and deployment type, and changes often, so check each vendor's model page for the exact model you plan to use.
Billing. OpenAI bills your OpenAI organisation directly. Azure and Bedrock charges appear on your existing Azure or AWS bill, which helps if you already have committed spend, startup credits or a reseller arrangement there. All three charge per token, with different rates for input and output, and the price per customer conversation is usually small but not zero. Our guide to LLM cost per customer reply shows how to estimate it.
How to choose and test in an afternoon
Benchmarks and marketing pages only go so far. A short hands-on trial with your own data tells you more. We have not run these tests ourselves; this is a method you can follow.
Write down your constraints
Required processing location, data retention limits, which cloud you already pay for, and the model family you need.
Shortlist one model per platform
Pick the closest equivalent on each platform that meets your constraints, using each vendor's model and region pages.
Prepare 20 real test prompts
Use anonymised examples of the actual task: customer emails, product questions, document summaries.
Run them and record results
Note answer quality, speed and token counts for each. Save everything in a spreadsheet.
Check the admin side
Confirm you can set retention, restrict API keys, see usage and set spending alerts on each platform.
Do it yourself or get help?
If you are a developer, getting an API key on any of the three takes minutes. The time goes into the parts that matter for a business: retention settings, keys stored safely, spending limits, logging and guard-rails on what the AI may answer.
Our Custom Tool or Dashboard service (from S$1,990, fixed quote after a free written scope) builds AI features into your own tools, deployed in your accounts with the source code yours. For a customer-facing assistant on your website, AI Website Assistant (S$790) is trained on your own content with guard-rails and hand-over to a human.
Frequently asked questions
Both say they do not train on API data by default. Microsoft states prompts and outputs are not available to OpenAI and are processed under Microsoft's data protection terms, with regional and Data Zone deployment options. OpenAI keeps API abuse monitoring logs for up to 30 days unless you are approved for zero retention. The better fit depends on your contracts and cloud set-up.
Partly, depending on the platform. As of October 2026, OpenAI's Singapore data residency stores data at rest in Singapore but does not process inference there, and needs approval. Amazon Bedrock has a Singapore Region endpoint. Azure has a Southeast Asia region, but Global deployments may process prompts wherever the model is deployed. Check the exact model and deployment type.
AWS says inputs and outputs are not shared with any model providers and are not used to improve base models. Some newer models require a retention mode where AWS keeps prompts and outputs within AWS for up to 30 days for human review by AWS, not by the model provider. You can choose zero retention, but that blocks models that require review.
It depends on the model, not the platform alone. All three charge per token with separate input and output rates, and OpenAI adds a 10% uplift for data residency on newer models. Compare the specific models you would use on each vendor's pricing page, and estimate your monthly volume before deciding.
No. All three can be used pay-as-you-go: an OpenAI API account, an Azure subscription or an AWS account. Some controls, such as OpenAI's zero data retention and data residency or Azure's modified abuse monitoring, require an application or a sales conversation, so plan for that if your use case needs them.