The god of post-training, actual measurement of the domestic large model GLM-5.3: 20 times more cost-effective than the top American AI Fable5

Kuai Technology reported on August 15 that Zhipu officially released the GLM-5.3 large model. Compared with the 5.2 version, the base model has not been upgraded. It still has 744 billion parameters, but the performance has been greatly improved, catching up with Fable 5. Domestic and foreign AI users are very surprised by the performance of GLM-5.3. Some even suspect that the GLM-5.3 is over-fitting the current evaluation projects. However, the actual test performance of netizens in the past few days has also denied the possibility of cheating in running scores. GLM-5.3 achieves such strong performance with 1/10 to 1/5 of the parameters of the large American model. It relies on their RL post-training. In this regard, GLM-5.3 is called the god of post-training, and the wisdom spectrum indicates that there is still potential to be tapped. It is estimated that the subsequent GLM-5.5 will be even stronger. This time GLM-5.3 not only catches up with the Fable level in performance, but also maintains its price advantage. Compared with GLM-5.2, the price has not increased. In addition, the amount of tokens used during use is reduced, and the overall cost is still reduced. Command Code used the FlappyBench test project to compare the performance and cost of the three models GLM-5.3, Fable 5 and GPT-5.6 Sol. The scores based on functionality, UX/UI and cost are as follows: Fable 5 → 9.5/10 · $0.420 GLM 5.3 → 9/10 · $0.018 GPT-5.6 Sol → 9/10 · $0.150 The result is obvious. Fable 5 wins in terms of quality, but the cost is 23 times that of GLM 5.3. In other words, GLM-5.3 is about 20 times more cost-effective. The output of GLM 5.3 is better than that of GPT 5.6 Sol, and the cost is only 1/8 of it, which is also a significant lead. Although they recommend that users choose Fable 5 if they value quality more, this is actually to show respect for Fable 5. The quality gap between GLM-5.3 and GLM-5.3 is not that obvious, but the 20 times price-performance ratio cannot be ignored. Whether it is an individual programmer or a company, the cost of long-term use is different.