Vibe coding, the most frustrating thing for people is
Every few days, a new model emerges that tops the charts, with data that is world-class.
You can't just try out every single one, so let's look at the bloggers on social media. You'll find that most of them are just hyping things up.
Then, after finally settling on a suitable model and product, you want to get down to some serious work. But it might not be long before you encounter a model that has degraded in performance.
It's really frustrating. Recently, I've noticed that Gemini 3 Pro and Claude 4.5 have both degraded in performance.
Yesterday, I used Antigravity's Gemini 3 to write a task, and the issues kept coming up. Alright, I tested a few commonly used models, and the results were:
Confirmed that Gemini 3 has degraded in performance.
Tested Claude Code using Claude 4.5, which also showed significant degradation.
Used GLM 4.7 in Claude Code, which recently claimed to be world-class, but although Claude 4.5 found more issues, it missed 1/3 of them.
In Cursor, GPT 5.2 performed the best, missing only one minor issue while finding all the others.
Based on this, GPT 5.2 is the most capable. The only headache is that the code written by GPT 5.2 is too simplistic, making it hard to understand, which is quite frustrating. 😀
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
