ChatGPT 5.x models max_tokens rename

[cberon]

OpenAI has changed the max_tokens parameter in 5.x models to be max_output_tokens (API) or max_completion_tokens (chat) as per https://help.openai.com/en/articles/5072518-controlling-the-length-of-openai-model-responses

Current fabro behavior still references max_tokens for openai and openaicompatible providers - this is causing issues with the newer models as not all hosting platforms support an alias for max_tokens. Using fabro with these providers (Azure as an example) requires custom integration or proxy rewrite.

Since gpt 4 was deprecated on Feb 13, 2026, update fabro default behavior to pass the value as max_output_tokens or max_completion_tokens. Consider adding a legacy setting in the config for maintaining compatibility with older 4.x models.

[swerner]

Thanks for raising this @cberon. Fabro's openai adapter already uses the Responses API and translates its max_tokens setting to max_output_tokens. If your Azure deployment supports /openai/v1/responses, you should be able to use it by configuring Azure as a custom provider with adapter = "openai" instead of openai_compatible.

The openai_compatible adapter uses Chat Completions and still sends max_tokens, so it doesn't currently support endpoitns requiring max_completion_tokens. Since compatible providers vary here, we'll need to work through how to make that selectable. If the Responses configuration doesn't work for your Azure deployment, would you be able to share a sanitized provider configuration and the error you receive?