When using Predicted Outputs, the number of tokens in the prediction that did
not appear in the completion. However, like reasoning tokens, these tokens are
still counted in the total completion tokens for purposes of billing, output,
and context window limits.