Response streaming returns a model's output as it is generated, in small pieces, instead of waiting for the complete response. It reduces the perceived delay before the first text is displayed, but complicates error handling and validation.
Example
An assistant's text appears word by word while the model generates it.