The inference accuracy of a single base model is often constrained by its reasoning capacity within an AI-model network. Although Token-level parallel collaboration with collaborative models can effectively improve inference accuracy, existing approaches generally rely on full-process Token-level parallel collaboration, resulting in substantial Token overhead and inference latency. To address these challenges, this study proposes a confidence-aware collaborative inference method, named CobCo. The proposed method leverages the base model’s confidence as a core control signal to regulate its inference behavior dynamically. When the base model has low confidence, collaborative inference is activated to reduce the risk of errors. Conversely, when confidence is sufficiently high, the base model switches to independent inference to reduce redundant collaborative interactions. To support this adaptive inference mechanism, a confidence evaluation algorithm tailored for Token-level parallel collaboration is designed. By treating collaborative outcomes as external references, the proposed method characterizes the decision quality of the base model during step-by-step inference and dynamically quantifies its confidence using a Beta distribution. Experimental results show that integrating CobCo improves the base model’s inference accuracy by 2.4% to 27.6% compared with independent inference. Compared with existing full-process Token-level parallel collaboration methods, CobCo significantly reduces Token overhead and inference latency while incurring only a small loss in inference accuracy. Notably, when the base model and collaborative models exhibit comparable performance, CobCo achieves up to a 20.8% reduction in Token overhead without any degradation in inference accuracy.