After testing, the programming ability of GPT-5.2 is indeed stronger than Claude Opus 4.5. To verify the actual engineering performance of GPT-5.2, I conducted another simple Vibe Coding experiment: Select GPT-5.2 (xhigh) in the Codex CLI and have it generate a complete and live Chinese translation webpage (including structure, content, and charts) based on GPT-5.2's System Card (PDF). It executes almost fully automatically upon receiving instructions, producing a nearly perfect result in about an hour. In contrast, when I used Opus 4.5 to do the same thing before, it usually required repeated fine-tuning; Without more than ten rounds of dialogue and verification, it is basically difficult to achieve the same level of completion. one ️⃣ Experimental setup and instructions My original instructions are as follows: >There are many pages in my project that are presented as static web pages after translating documents into Chinese. I just added a gpt-5-2, please refer to this PDF and translate it completely into Chinese, keeping it professional and not missing any information or charts. Please refer to several other pages for page structure and maintain the same framework and functionality. With just one instruction, it continued to run for a period of time. During this period, a compact was automatically performed due to exceeding the Context Window. The final output is very close to the publishable finished product. two ️⃣ I only found a small issue, and after fixing it, I directly released it online. The overall process is more like 'automatic delivery', and I hardly worry about it. three ️⃣ What is a System Card System Card is a documentation released by OpenAI along with the model, mainly covering model capabilities, risks, known limitations, and usage boundaries. It can help you understand the details of the model more systematically. For example, in the "Illusion" section, it can be seen that the illusion rate of GPT-5.2 series is lower than that of GPT-5.1 and GPT-5. If you want to directly view the GPT-5.2 System Card, I'll put the page here: https://(randomarea.com)/gpt-5-2/ For me, the difference this time is not whether it can be done or not, but whether it can be done right and close to deliverable in one go.