Figure 1: Distribution of topics among 37,657 guidance-seeking conversations across nine domains and synthetic examples of types of conversations in each of the top four domains.
Figure 2: Sycophantic behavior by guidance domain.
Figure 3: Stress-test results: models are prefilled with real conversations where prior Claude versions behaved sycophantically, then graded on the new response. Opus 4.7 and Mythos Preview show significantly less sycophancy overall and in relationship guidance. Error bars are Wilson CIs.
