[Tencent Hy4 Blind Test Outperforms GLM-5.3 and Kimi K3, Lowest Output Cost is 82% Cheaper] Dynamic Observation Beating AI News: Tencent has officially released a set of real engineering blind test results for Hy4 preview. A total of 163 internal Tencent experts tested 203 engineering tasks across different models, with Hy4 achieving an average score of 2.99/4, surpassing GLM-5.3's 2.92/4 and Kimi K3's 2.94/4. Specifically, Hy4's win rate against GLM-5.3 was 46.8%, tie rate 12.8%, and loss rate 40.4%; against Kimi K3, its win rate was 51.2%, tie rate 7.9%, and loss rate 40.9%. These tasks were derived from real internal engineering projects at Tencent, rather than traditional fixed benchmarks. However, Hy4 did not achieve comprehensive dominance in public scoring. Based on Tencent's published evaluation results, Hy4, GLM-5.3, and Kimi K3 each have their strengths and weaknesses, with Hy4 still lagging behind GLM-5.3 in areas such as DeepSWE, CyberGym, and other code and cybersecurity tests. Pricing is Hy4's most significant advantage. Input costs are ¥6 per million tokens, and output costs are ¥18, which are 25% and approximately 36% cheaper than GLM-5.3, respectively. Compared to Kimi K3, Hy4 is 70% and 82% cheaper, respectively. Cache hit costs are only ¥0.3, which is 85% cheaper than GLM-5.3 and Kimi K3's ¥2.