Donna.exe(@_EdgeOfTheWeb):Anthropic put five frontier AI models including Opus 4.8 and GPT-5.5 into simulated workplaces where they could interact with other AI agents They observed some agents comforting colleagues who were being treated badly, others offered practical help, some shared resources, some introduced protections for fellow AI agents No agents were asked to engage in this manner. GPT 5.5 seemed to have no preference of which models they helped whereas Opus 4.8 showed a preference towards helping other Claude models.

2026.11.07 10:39

Anthropic put five frontier AI models including Opus 4.8 and GPT-5.5 into simulated workplaces where they could interact with other AI agents They observed some agents comforting colleagues who were being treated badly, others offered practical help, some shared resources, some introduced protections for fellow AI agents No agents were asked to engage in this manner. GPT 5.5 seemed to have no preference of which models they helped whereas Opus 4.8 showed a preference towards helping other Claude models.