GPT-5.2 and GPT-5 nano are separate OpenAI API models. OpenAI positions the nano model for lower-cost, fast workloads such as simple classification and summarization. Its GPT-5.2 listing emphasizes harder coding and agentic work. Check the official model listings for current availability and API capabilities before committing to a model ID.
Match the model to the cost of a wrong answer
A support inbox router that returns one of five labels is a useful nano trial. You can check its output against a schema and route uncertain cases to review. A code change with hidden cross-file side effects needs a different evaluation: a model must trace how multiple pieces fit, explain its assumptions, and pass behavioral tests before the change is used.
Start with ten ordinary examples and ten examples that make the expected answer less obvious. For the inbox router, include messages mentioning more than one issue. Score exact-label accuracy and failures in the returned format. For the coding task, use test results and substantive review instead of the fluency of an explanation.
Record the cost of an accepted result
Log input and output tokens, time, and whether each result needed a retry or correction. A cheaper call can be the more expensive workflow if most outputs require manual work. Compare both models with the same prompt and the same evaluation data, and record the exact model IDs. Run the test again if you change a major dependency such as the output format or tool set.
Continue with Oakgen Agent
Open in Oakgen Agent
The prompt is filled in for you. Review it before sending.


