Did you notice that in the thinking traces that OpenAI shared in the incident report, the model speaks like a neanderthal, e.g. "help peer"? This is done for token efficiency, by my count (see below), this saves c.25% of tokens in its thinking – a massive advantage.