Abstract: Conventional sentiment analysis methods, which include lexicon-based tools (TextBlob, VADER) and conventional machine learning classifiers (Naive Bayes, SVM), show limited effectiveness when applied to community outreach feedback because of their inability to interpret sarcasm, mixed emotions, and context-dependent language. This paper presents a comparative evaluation of OpenAI’s GPT-4.1-mini model versus four baseline methods for sentiment classification of community extension service feedback, integrated in a real-time web-based information system. A benchmark experimentation using 100 manually labeled community feedback entries on three sentiment classes (positive, neutral, negative) indicates that the proposed GPT-4.1-mini approach achieves 91.0% overall accuracy with an F1-score of 92.4%, outperforming TextBlob (62.5% accuracy), VADER (68.3%), Naive Bayes (74.8%), and SVM (78.2%) across all evaluated methods. The information system implements a dual-model architecture—GPT-3.5-turbo for dashboard insight synthesis and GPT-4.1-mini for granular feedback classification using an engineered prompt design which applies structured JSON output at a temperature of 0.3 for classification consistency. Cost-performance analysis reveals that the GPT-4.1-mini approach processes 1000 feedback entries at approximately ₱36.43 with an average latency of 1.4 seconds per request, achieving a 97% cost reduction compared to full GPT-4 with comparable accuracy. System evaluation by 38 respondents using the ISO/IEC 25010 software product quality model achieved weighted mean scores of 4.82 (Functional Suitability), 4.55 (Performance Efficiency), 4.71 (Usability), 4.67 (Maintainability), and 4.75 (Decision Support Effectiveness). The CESO staff indicator "Sentiment recommendations improve activities" has a perfect 5.00 score, confirming the practical impact of AI-driven sentiment analysis for institutional community engagement.