Despite rapid progress in AI capabilities, there is currently no benchmark designed to evaluate how well an AI functions as a cooperative social partner over time. Existing benchmarks focus on task accuracy, short-horizon preference alignment, or isolated safety behaviors, but largely ignore relational properties humans care about, such as consistency, reciprocity, responsiveness, fairness, and repair after missteps. Although some multi-agent and social dilemma benchmarks probe coordination or incentive alignment, they do not evaluate AI behavior from the perspective of a human engaged in an ongoing relationship. There is a need for systematic ways of assessing whether AIs support trust, cooperation, and mutual benefit in the kinds of repeated interactions where real social impact occurs. In other words, how good are AIs at being a "true friend" to us?