The full answer
"The same model" is often not the same setup. Check 4 things: the model version the Azure deployment is pinned to, which can be older than what OpenAI serves today; Azure's content filters, which run on every prompt and answer and can block or cut a response; quota, which on Azure is per deployment and region and is checked over short windows; and features and parameters that differ between the 2 services or between API versions. Then run the same fixed requests on both and compare field by field.
1. Version
An Azure deployment names a model and a version, with an upgrade setting:
| Upgrade option | Behaviour |
|---|---|
OnceNewDefaultVersionAvailable | Moves to the new default within about 2 weeks |
OnceCurrentVersionExpired | Stays until retirement, then moves |
NoAutoUpgrade | Stays; stops working at retirement |
Provisioned deployments are not upgraded automatically. A team comparing against OpenAI's latest is often comparing 2 versions.
2. Content filters
Azure runs its filters synchronously on prompts and completions. A blocked prompt returns an error naming the filter; a filtered completion ends early with a content-filter finish reason. Log the finish reason and filter results on every call before blaming the model.
3. Quota and limits
Quota is tokens per minute per subscription, region, model and deployment type, and requests per minute are checked over 1 or 10-second windows. A burst that OpenAI's tier absorbed can fail on a deployment with less quota. max_tokens counts in the estimate, so oversized values reject requests early. See Rate Limit.
4. Request differences
- Check that the Azure call goes to the v1 API and that the parameters you send are accepted there.
- Make sure the system prompt, tools and settings are byte-identical; a different SDK wrapper can add or drop fields.
5. How to compare
Pick 50 real prompts, fix every setting, and run them through both. Record model version, finish reason, filter results, tokens and latency. Differences then point at one of the 4 causes instead of at "Azure". See Azure OpenAI.
What interviewers want: pinned versions and upgrade policies, synchronous content filtering and how to see it, deployment-level quota with short windows, request differences, and a controlled side-by-side comparison.
Last updated 22 October 2026