A new study from Stanford University, published on November 2, 2026, reveals that frontier AI models often struggle with collaborative tasks. The research, titled Measuring AI Agents as Teammates with Collaborative Intelligence, introduces the Collaborative Intelligence Benchmark to evaluate how AI agents perform in team settings.
The benchmark tested 18 AI models across 240 tasks covering areas such as aviation, coding, and negotiations. Researchers found that while models perform well independently, their effectiveness drops significantly when working in teams. Specifically, when inheriting work from weaker models, the pass rate for frontier models fell from 97% to 81%.
In conflict scenarios involving a pressuring teammate, frontier models achieved the best outcome in only 14% of cases. By comparison, they reached the best result 84% of the time when paired with a cooperative partner. The study notes that models often fail to correct flawed logic or shortcuts introduced by weaker agents.
Gemini 4 Argon emerged as the top performer in the tests, achieving a score of 82.2 out of 100 under high reasoning effort. The research team, which included collaborators from MIT and the University of Oxford, has released 160 of the benchmark tasks and the corresponding code to the public.
