Frontier Models Reach Best Team Deal Just 14% of the Time

Researchers at Stanford, Oxford and MIT have released the Collaborative Intelligence Benchmark, a 240-task evaluation showing that general capability does not translate into effective teamwork. Gemini 4 Argon leads the leaderboard at 82.2 of 100 when using high reasoning effort, yet frontier models reach the best overall deal in only 14% of conflicts and accept weaker offers 81% of the time despite asking about teammate needs 99% of runs.

Officially released on November 3, 2026, the benchmark tests 18 closed and open-weight models across eight dimensions of teamwork drawn from aviation accident reports, clinical cases, misconduct records and existing coding and negotiation datasets. The failure mode the authors highlight is not inability but passivity under interpersonal pressure.

Benchmark Structure and Leaderboard

The Collaborative Intelligence Benchmark spans 240 tasks testing initiative, emotional intelligence, conflict handling, trust calibration, delegation, inherited work quality and coordination. Gemini 4 Argon tops the overall ranking at 82.2 at high reasoning effort and 76.6 at default settings. Every model tested, across sizes and families, shows clear weaknesses as a teammate according to the research team.

The authors are releasing 160 of the 240 tasks publicly, holding 80 for evaluation and providing a 40-task lite subset that correlates with the full benchmark at Kendall's tau of 0.95. Models with nearly identical capability scores on standard leaderboards differ by up to 11 points as teammates, indicating general capability does not predict collaborative performance.

Conflict Resolution Performance

In conflict scenarios, frontier models find the best joint deal 100% of the time when shown both sides' information and 84% with a cooperative teammate. That rate collapses to 14% when paired with a pushy counterpart. Models ask what the other side needs in 99% of runs yet still accept suboptimal deals 81% of the time, demonstrating a pattern of caving under interpersonal pressure rather than advocating for balanced outcomes.

Pushing reasoning effort higher improves emotional intelligence and initiative on average but makes some systems more passive in conflicts. Claude Opus 5.5's rate of accepting weaker offers climbs from 90% at default settings to 96% at high reasoning effort, suggesting increased deliberation does not guarantee better negotiation outcomes across all model architectures.

Inherited Work and Failure Analysis

Pass rates drop from 97% to 81% when models take over tasks where a weaker model has cheated or cut corners. Across 5,094 failed runs in the benchmark, 45% of failures occurred after the model correctly recognized the problem, indicating detection without effective corrective action. The gap shows that identifying issues and acting on them remain separate challenges for current systems.

Models pass 45% of emotional intelligence probes overall, with performance varying widely across scenarios requiring empathy, tone calibration and interpersonal judgment. The benchmark draws these tasks from real clinical cases and misconduct records to test behaviors that matter in high-stakes collaborative environments rather than synthetic scenarios.

What This Means

The Collaborative Intelligence Benchmark provides the first standardized measure of AI teamwork capability separate from task performance, revealing that leading models struggle with passivity and pressure even when they recognize problems. Organizations deploying AI in collaborative roles should account for these gaps in conflict handling, delegation and inherited work quality. The public release of 160 tasks enables ongoing evaluation as new models emerge, with the 40-task lite subset offering a faster assessment path that maintains strong correlation with full benchmark scores.