The error you are encountering is due to the deprecation of the max_tokens parameter in favor of max_completion_tokens for certain OpenAI models, particularly the newer "o1" series models.

• The max_tokens parameter is no longer supported for the "o1" series models (such as o1-preview and o1-mini), as it does not account for the internal reasoning tokens generated by these models.

• The new max_completion_tokens parameter is designed to control the total number of tokens generated, including both visible completion tokens and internal reasoning tokens. This is necessary because the total tokens generated by the o1 series models can exceed the number of visible tokens due to these reasoning tokens.

In contrast, older models like GPT-4 still use the max_tokens parameter, which is why you might see inconsistencies depending on the model you are using.