Predicting your own family versus everyone else

Pooled accuracy of hard predictions, split by whether the opponent came from the same provider. Bars are 95% intervals from 5,000 whole-game bootstrap resamples.

Opponent from the same providerOpponent from another provider
30%40%50%60%70%80%90%100%OpenAI → own family36/55OpenAI → others78/145Anthropic → own family17/24Anthropic → others31/62Correct predictions (% of scorable)

Within minus outside: OpenAI +11.7 pp (95% CI -1.6 to 25.7), Anthropic +20.8 pp (0.0 to 44.2). Open-weight models come from different labs, so they have no family to predict within.