Yoav Goldberg Flags Millions of Reasoning Tokens per Game

Yoav Goldberg posted that the cheapest run of a setup used around six to seven million tokens per game, calling the volume of reasoning tokens substantial. He suggested the model was likely trained directly in that environment. Goldberg also noted interest in comparing no-reasoning versus low-reasoning performance on the standard harness. The post includes a table image with columns for reasoning effort and harness types.