Why iFlytek Spark X2 Scored 708 on the Gaokao: An In‑Depth Model Analysis

Recent Gaokao results for AI "exam‑takers" were released, featuring a blind assessment by two experienced teachers of eight domestic and foreign large models. Eight AI candidates answered full‑subject questions, and scores were tallied separately for the history‑track and physics‑track categories.

Overall, Claude Opus 4.8 and iFlytek Spark X2 both achieved 708 points in the physics‑track, while Spark X2 alone surpassed 700 points in the history‑track, reaching the "Guangdong shielded‑student" level. The modest gaps between top models indicate that total ranking depends more on stable performance across subjects than on isolated peaks.

ChatGPT 5.5 Pro and Claude Opus excel in long‑text generation and argumentative essays, yet they fall behind in the history‑track, illustrating that even models with strong language abilities exhibit tiered differences in balanced performance. Spark X2 leads both tracks, which the evaluation report attributes to its equilibrium across language comprehension, mathematical reasoning, and integrated analysis, without any single subject pulling the score.

Subject‑specific tests reveal finer details: In a math paper from XinJing Bao, Spark X2 scored 148 points, followed by Kimi (145), DeepSeek (144), Zhipu (143), MiniMax (142), and ChatGPT (137). In a Shanghai‑track essay contest organized by 澎湃新闻, Spark X2 earned 65.5 points, ahead of Gemini (64.5) and Doubao (64). For English essays, 观察者网 placed Spark X2 and ChatGPT 5.5 Pro in the top tier.

The report explains Spark X2's superiority by highlighting its adherence to process compliance. In mathematics, the model provides complete, textbook‑consistent derivations, avoids step‑skipping, and even offers dual solution paths, demonstrating a "number‑shape" reasoning advantage. Similar consistency in physics, chemistry, and biology reduces loss points, while its essay scores benefit from balanced module performance and strong logical structure.

Underlying these results is the quality of training data. General LLMs rely on publicly available internet text, which lacks granular educational data such as step‑by‑step student solutions, error type distributions, and teacher annotations. iFlytek has accumulated high‑density, professionally curated educational data over 22 years, including alignment data between machine assessments and human scores collected since 2012. This data feeds a "teaching‑thought‑chain" that structures teacher judgment logic into trainable formats, enabling the model to learn evaluation standards directly.

While high scores in a one‑off, highly structured test demonstrate a model's capability, they do not guarantee success in everyday classroom settings, which involve continuous, context‑rich interactions, varied teacher habits, and diverse school infrastructures. The real barrier lies in deployment: technology must be usable, teachers must be willing to adopt it, and schools must sustain the operational environment.

To address these deployment challenges, iFlytek pursues a hardware‑software integrated approach. By developing proprietary terminals that control data entry and usage environments, the company embeds the model into three core teaching scenarios—classroom interaction, post‑class assignment grading, and at‑home tutoring—creating a closed data loop: usage generates labeled data, which refines the model, which in turn expands product deployment.

From an industry perspective, AI education is moving from the first stage—proving models can answer questions correctly—to a second stage where products must embed seamlessly into teaching workflows and continuously harvest valuable scene data. This shift raises the competitive bar beyond model size to long‑term, deep domain expertise and data acquisition.

Ultimately, the article argues that AI's high‑score achievements should be viewed not as a threat to teachers but as a means to democratize quality education, offering personalized support to every learner and reducing the historic concentration of educational resources.