Black-box distillation trains student models using only teacher-generated outputs, making it practical when teacher models are accessible only through APIs. However, existing methods either rely mainly on positive imitation or require costly full-response rollouts to obtain corrective signals. We show that such corrective signals can be obtained locally from teacher trajectories alone. At each position, the teacher token provides a positive anchor, while a high-probability alternative sampled from the student provides an informative negative. Based on this insight, we propose Token-level On-policy contrastive Learning for blackbox Distillation (TOLD), which samples a single negative token from the student’s distribution and optimizes a Bradley-Terry objective anchored by negative log-likelihood. This enables the student to reinforce teacher tokens while suppressing its own likely mistakes, without teacher logits, reward models, or full-response rollouts. Experiments on reasoning and instruction-following tasks show that TOLD outperforms or matches strong black-box distillation baselines and achieves up to 1.88 times training speedup over rollout-based on-policy baselines.