Background and Aim: Accurate preoperative identification of clinically significant prostate cancer (csPCa) remains challenging due to the limited specificity of prostate-specific antigen (PSA) and the experience-dependent inter-reader variability of PI-RADS scoring. Large language models (LLMs) have shown potential in medical decision support, but their comparative performance against radiologists of varying experience and their independent predictive value in csPCa diagnosis are poorly understood. Methods: This multicenter retrospective cohort study included 1,709 patients who underwent prostate biopsy across three tertiary hospitals. Four LLMs (DeepSeek-V4 Flash, GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Flash) generated csPCa probabilities and PI-RADS scores from anonymized MRI reports combined with clinical variables. Three radiologists (junior, intermediate, senior) independently assigned PI-RADS v2.1 scores. Diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, calibration metrics, and decision curve analysis. Logistic regression and incremental value analyses were performed. Subgroup analysis was conducted in the PSA gray zone (tPSA 4–10 μg/L). Results: Claude Opus 4.6 achieved the highest AUC among LLMs (0.840, 95% CI: 0.821–0.859), comparable to GPT-5.4 (0.839) and Gemini 3.1 Flash (0.835), and significantly outperformed DeepSeek-V4 Flash (0.777, P < 0.001). Claude Opus 4.6 showed similar discriminative ability to the senior radiologist (AUC 0.840 vs. 0.842, P = 0.866) and significantly outperformed junior (0.704) and intermediate (0.747) radiologists. PI-RADS agreement was fair with the senior radiologist (κ = 0.383) but slight with junior/intermediate radiologists. In multivariable analysis, adding Claude Opus 4.6 probability to clinical variables improved AUC from 0.851 to 0.867, but after including the senior radiologist’s PI-RADS score, the LLM probability provided no incremental value (ΔAUC = −0.0002, P = 0.616). In the PSA gray zone (n = 803, csPCa prevalence 27.0%), all evaluators showed decreased performance; the best LLM (Gemini 3.1 Flash, AUC = 0.755) performed comparably to intermediate radiologists but remained inferior to the senior radiologist (0.826). Conclusion: Claude Opus 4.6, GPT-5.4, and Gemini 3.1 Flash achieved diagnostic accuracy for csPCa comparable to that of a senior radiologist and significantly better than junior and intermediate radiologists. However, LLMs probability did not provide incremental predictive value beyond the senior radiologist’s PI-RADS score. In the challenging PSA gray zone, LLM performance dropped to the level of an intermediate radiologist. LLMs may serve as useful decision-support tools, especially in resource-limited settings, but are unlikely to replace expert human interpretation.