OpenAI vs Anthropic
Model quality moves every few months. Where the data may be processed, and what your compliance team can sign, moves far more slowly.
The short answer
Choose Anthropic when a contract term about training data has to survive legal review, and choose OpenAI when you need the widest spread of model types behind one account, including images, speech and live audio.
Both vendors say your API data does not train their models. They say it differently, and the difference matters to a lawyer. Anthropic's commercial terms state that it "may not train models on Customer Content from Services". OpenAI documents training as opt-in, so the protection is a default setting rather than a contractual bar.
For most builds the honest answer is that either will work. Write your code so the provider can be swapped, then choose on procurement rather than on a benchmark you cannot reproduce.
Side by side
Taken from each vendor's own documentation and checked in August 2026. Model rankings are left out on purpose: they change faster than this page can.
| Dimension | OpenAI | Anthropic |
|---|---|---|
| Training on your API data | Not used unless you opt in to share it | Commercial terms say it may not train on it |
| Default log retention | Abuse logs up to 30 days, longer if law requires | Not stated as a number in the page we checked |
| Zero retention | Eligible customers may apply, subject to approval | An approved arrangement for eligible customers |
| Health data | BAA cover is endpoint by endpoint, with conditions | HIPAA-ready setup, BAA available and switched on by you |
| Where it runs | OpenAI's own API, plus Microsoft Foundry | Own API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry |
| Breadth of model types | Text, image, speech, transcription, live audio | Text models across several sizes |
| Cheap bulk work | Batch API, jobs complete within 24 hours | Not covered in the pages we checked |
| Buying signal | You want one vendor for several media types | You want the strongest written data term |
- PromptsTuned per model. Expect rewriting
- Tool callsDifferent shapes, same idea
- Eval setThe only thing proving the swap worked
- Cost modelToken counting differs per vendor
- ContractThe slowest step, and the first to start
Teams budget for the code and forget the eval set. The eval set is what makes a switch a decision rather than a hope.
OpenAI
Where it wins
- One account covers text, images, transcription and live audio, so a voice feature does not mean a second vendor.
- The Batch API runs large offline jobs at a documented discount, with each batch completing within 24 hours.
- The published compliance list is long, including SOC 2 Type 2, ISO 27001 and ISO 42001.
- Regional processing is documented for the United States, Europe and the United Arab Emirates, with storage-only regions elsewhere.
Where it hurts
- The training protection is a default you must not turn off, rather than a promise written into commercial terms.
- Zero retention needs approval, and the docs say the controls are subject to prior approval by OpenAI.
- Health data cover is granular. Web search with live internet access is documented as not HIPAA eligible and not covered by a BAA.
- Some endpoints keep data until you delete it, which is a different rule from the 30-day one people quote.
Anthropic
Where it wins
- The commercial terms bar training on customer content, which is a stronger artefact than a settings page in a review.
- Claude runs on Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry, so it can sit inside a cloud contract you already signed.
- ISO 42001 for AI management sits alongside the usual SOC 2 and ISO 27001.
- Anthropic documents the failure mode of long prompts by name, calling it context rot, rather than selling context size as free capability.
Where it hurts
- The range is narrower. Image generation, speech and live audio are not part of the offer, so those features need another vendor.
- Zero retention is still an approved arrangement, and safety classifier results are kept even under it.
- A default retention period in days was not stated on the page we checked, so a security questionnaire may need a direct answer from sales.
- HIPAA settings are switched on by an owner rather than applying automatically, which is easy to miss during setup.
How to choose
- Choose Anthropic if a data protection officer must sign the arrangement, because a term in commercial terms is easier to defend than a default setting.
- Choose Anthropic if your company already buys through AWS or Google Cloud and would rather add a model than add a supplier.
- Choose OpenAI if the product needs images, transcription or a live voice interface alongside text.
- Choose OpenAI if you have a large offline job, such as classifying a back catalogue, where a 24-hour turnaround is fine.
- Choose both if you can afford the eval work. Routing different jobs to different providers is normal, and it removes a single point of failure.
- Choose neither yet if nobody has written down what a good answer looks like. Without a scored test set you cannot tell which vendor is better for you, and you will pick on vibes.
Which provider fits your build?
Answer about your constraints, not about which model impressed you last week.
What does the product need to handle?
Who has to approve the vendor?
What cloud contract do you already hold?
How do you currently judge answer quality?
Every outcome
- Anthropic
- Your buying is governed by legal review or an existing cloud contract, and the work is text.
- OpenAI
- The product spans several media types, or you have bulk offline work that can wait a day.
- Build the test set first
- Without scored examples, any provider choice is a guess you cannot defend or revisit.
Systems we have shipped on these models
Questions buyers actually ask
01Should we commit to one provider or stay portable?
Stay portable at the boundary, and commit everywhere else. Keep model calls behind one internal interface so a swap touches a few files rather than the whole codebase. Full portability is a myth worth abandoning early: prompts are tuned per model, and the tuning does not transfer even when the code does.
02Which vendor is safer for regulated data?
Neither is safe by default, and both offer stronger arrangements you have to ask for. Anthropic publishes a HIPAA-ready configuration with a BAA available, activated by the account owner rather than applying automatically. OpenAI's BAA cover varies by endpoint, and it documents that web search with live internet access is not covered. Read the endpoint list before designing the feature, not after.
03Does running Claude on AWS change who sees our data?
It changes which contract governs it, which is usually the point. Anthropic documents Claude on Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. Buying through a cloud you already use means your existing agreement and billing apply, and your security team reviews a supplier they have already approved.
04How much work is it to move from one to the other?
The code is the easy half. Expect to rewrite prompts, adjust tool definitions, and re-tune anything that depended on a particular model's habits. Budget most of the effort for re-running your evaluation set, since that is the only evidence the switch did not quietly make things worse.
05Is the bigger context window a reason to pick one?
Rarely, and Anthropic itself is blunt about why. Its docs state that as token count grows, accuracy and recall degrade, and it names the effect context rot. Treat a large window as room to manoeuvre rather than as a licence to paste everything in.
