Summarizing often fails if not using davinci

Using, for example, Playground (but also with the API), summarization fails (“The model predicted a sequence…”) even when using exactly the same text - if I am not using davinci. The 2nd grader Jupiter example will fail using curie if I attempt the summarization 2/3/4 times (and carefully deleting the summary text underneath).

Is this expected behaviour? Or am I supposed to clear something in between summaries (‘Regenerate’ button does nothing)?

This behaviour also occurs when using the API and I get empty text returned.