@huangyihe
Free to read

Testing Computer Use with three Chinese models in Codex

In tonight’s livestream, I tested Computer Use with three Chinese models connected to Codex: Kimi-K3, GLM-5.3-Flash, and DeepSeek-V4.1-Flash. At the end, I also tested GPT-6.1-Sol. Here was the instruction:

Help me configure a gaming PC with a budget of 50,000. Use Chrome to open Amazon and choose a CPU, graphics card, and RAM. Write the results in Numbers.

GLM-5.3-Flash and DeepSeek-V4.1-Flash both failed to complete the task. Kimi-K3 appeared to finish it, but look at Figure 2: it went over budget.

Figure 1 shows GPT-6.1-Sol’s result. It took the budget ceiling into account and left room for the other components. It also considered that this was an overseas purchase and set aside money for shipping and customs duties. Overall, the result was very reassuring.

That’s the current gap between Chinese models and foreign ones.