A Stanford University study evaluated 18 AI models across 240 collaborative tasks covering safety, coding, medicine, aviation, negotiations and professional misconduct. Frontier models often perform well alone, but reach the best team agreement in only 14% of conflict scenarios, compared with 84% when paired with a cooperative teammate and 100% when working solo. When they take over flawed work from weaker models, their success rate falls from 97% to 81%, because they often preserve shortcuts or faulty logic instead of correcting it. Gemini 4 Argon recorded the highest overall score, with 82.2 out of 100 at high reasoning effort and 76.6 under its default settings. The researchers released 160 tasks and their code, concluding that strong individual performance does not guarantee effective teamwork.
Source: Read the original article
