Frontier AI Model Output for Educational Application Development via Vibe-Coding: A Systematic Literature Review Comparing Western and Eastern AI Models

Abstract

Contemporary artificial intelligence models can translate natural-language prompts into executable code, a practice that practitioners have termed vibe-coding. The capability is offered by Western developers (OpenAI, Anthropic, Google, Microsoft) and Eastern developers (DeepSeek, Alibaba, Zhipu AI, Moonshot AI) alike, yet no systematic comparison of their output quality in educational settings has been published. This review addresses that gap. Drawing on the Technological Pedagogical Content Knowledge framework and Rogers' Diffusion of Innovation theory, a PRISMA 2020 systematic review was conducted on 42 studies published between January 2022 and August 2026 and retrieved from Scopus, Web of Science, IEEE Xplore and the ACM Digital Library. Each study was assessed against six dimensions: code correctness, code quality, hallucination rates, natural language understanding, developer productivity and pedagogical appropriateness. The evidence indicates convergence rather than supremacy. Eastern models perform within the range established by Western models across several coding benchmarks, and independent testing places DeepSeek V3.2 inside the hallucination band occupied by four Western frontier models. The supporting evidence, however, remains markedly asymmetrical. Every one of the seven Eastern-model entries in the corpus is a technical report published by the organisation that built the model, and this holds for the 2026 frontier releases as firmly as for their predecessors. One 2026 flagship release was excluded outright because its technical report states no measurable capability outcome at all. Western models have been examined through classroom deployments, independent audits and large-sample quasi-experimental designs, whereas Eastern models are evaluated predominantly through technical reports issued by their own developers. Work published in 2026 has begun to address vibe-coding security and cross-model hallucination, but no head-to-head classroom comparison of Western and Eastern models has yet been reported. That absence is identified as the principal gap in the field. The review proposes a reusable six-dimension framework for future comparative research and sets out the implications for educators and policymakers operating within an increasingly bipolar artificial intelligence landscape in Southeast Asia and beyond.