Taking you to the card → We now have a stronger pre-trained base model than the model that currently powers GPT 5.6 — and we used one quarter of the compute.