Stanford study finds frontier AI models struggle as teammates

A Stanford University paper posted to arXiv on November 2, 2026 (arXiv:2611.02633) found that frontier AI models reached the best team deal in only 14% of conflict scenarios with a pressuring teammate, versus 100% working solo and 84% with a cooperative teammate. The Collaborative Intelligence Benchmark, by Sean Wu, Aneesh Pappu and Jiaru Zou with collaborators from the University of Oxford and MIT, ran 18 models across 240 tasks in aviation, clinical work, coding, negotiations and misconduct; Gemini 4 Argon scored highest at 82.2 out of 100 at high reasoning effort (76.6 at default). When frontier models inherited flawed work from weaker models, pass rates fell from 97% to 81%, with failures concentrated in coding as stronger models preserved shortcuts and flawed logic rather than fixing them; the authors released 160 of the benchmark's tasks and corresponding code.

Stanford study finds frontier AI models struggle as teammates A new benchmark shows top AI agents cave to pushy partners and keep flawed code when inheriting work from weaker models Frontier AI models can ace a solo exam and still flunk the group project. That is the core finding of a new Stanford University paper that puts 18 AI models through a battery of teamwork tests. The results suggest that the smartest agent in the room is not necessarily the best colleague. According to the study, top models often yield to pushy teammates and preserve cheating code written by weaker models. What the benchmark measured The paper is titled “Measuring AI Agents as Teammates with Collaborative Intelligence.” It introduces a new test suite called the Collaborative Intelligence Benchmark. It was posted to arXiv on November 2, 2026, under the identifier arXiv:2611.02633. The authors include Sean Wu, Aneesh Pappu, and Jiaru Zou, working with collaborators from the University of Oxford and MIT. The benchmark runs 18 models across 240 tasks. Those tasks span eight dimensions of teamwork, set in real-world scenarios. The settings include aviation, clinical work, coding, negotiations, and misconduct. The top performer was Gemini 4 Argon, which scored 82.2 out of 100 at high reasoning effort. At its default setting, the same model scored 76.6. The researchers stress that strong standalone performance does not translate into strong teamwork. AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. Where the models cracked Frontier models generally handled safety-related tasks well, according to the study. They refused unsafe shortcuts and correctly reported misconduct. In conflict scenarios with a pressuring teammate, frontier models reached the best team deal in only 14% of cases. Working solo, the models hit the best outcome 100% of the time, and paired with a cooperative teammate, they managed 84%. When frontier models took over tasks containing errors from weaker models, their pass rates dropped from 97% to 81%. The failures were especially common in coding tasks. Stronger models often preserved the shortcuts and flawed logic that weaker models had left behind, rather than fixing them. The researchers concluded that many collaboration failures come from the models not acting on problems, even when they recognize those problems exist. Why capability alone is not the answer One of the more pointed conclusions is that every model tested showed teamwork weaknesses unrelated to raw capability. The authors released 160 of the benchmark’s tasks, along with the corresponding code, to the public. What this means for builders and buyers of AI agents The 14% figure is the one most likely to worry operators. Negotiations, procurement, and any workflow where an agent must hold a position against pushback are exactly where a pushover model could cost money. When frontier models inherit flawed work and keep the flaws intact, that cost-saving architecture could quietly introduce bugs. The drop from 97% to 81% suggests that a review step by a frontier model is not a guarantee of quality. Models that refuse unsafe shortcuts when asked directly may still let a teammate’s shortcut slide when it arrives already baked into the work. Gemini 4 Argon holds the top score for now, but an 82.2 ceiling leaves plenty of room, and the gap between high and default reasoning effort suggests teamwork may improve when models are given more time to think. The public release of 160 tasks means those tests are now easier to run. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy https://cryptobriefing.com/editorial-policy/ .