points
How representative that is of real world usage, I don't know.
In their benchmark GPT 5.6 Sol performs suspiciously poorly compared to the former models.