My Subjective Impressions of Opus 4.6, GPT 5.4 xhigh, and Gemini 3.1
Translated from the 中文 original.
Day to day I use Gemini 3.1 and GPT 5.4 thinking; for coding it’s mostly GPT 5.4 thinking and opus 4.6. My access is mainly the Gemini app and antigravity and the API from gemini one pro, plus codex and the app from chatgpt teams. Everything I use is the official direct connection — no community proxy sites.
Opus 4.6’s quota is too small, so I don’t use it much, but there’s certain work that seemingly can only be done well by opus 4.6, which is impressive… Lately I’ve been vibe-coding some projects. Opus 4.6 is genuinely great at organizing design docs, doing requirements design, and so on…
Opus 4.6 is very good at writing docs, doing requirements and feature design, and in terms of feel its attention is strong. Hand it a very long requirements doc and it organizes it well, understanding every detail of the document. Code… my antigravity quota isn’t that big… so I don’t have deep experience with it, but my impression is that its code has a few more bugs than gpt5.4 xhigh.
GPT 5.4 xhigh is very good at writing backend code and algorithmic logic, extremely good at doing codex reviews — GPT5 is arguably the first model that made me feel reliable on code. But the frontend this thing produces is pretty questionable… not just ugly, the UX design is also a total mess, and you can’t figure out how to use what it makes.
Its docs aren’t great — doesn’t talk like a human, likes to write things nobody can understand, and judging by the results, I’d guess it can’t understand what it wrote either ().
Also, in long documents, gpt 5.4 xhigh struggles to reason through certain trade-offs.
Gemini 3.1 isn’t suited for writing code or organizing docs; this generation’s attention is too poor — make the requirements doc a bit more detailed and it starts dropping things left and right. But it’s great for learning. Even though overall it’s a step down compared to 2.5 pro, you can still squeeze a lot of knowledge out of Gemini’s mouth, while it won’t fail to talk like a human the way GPT does.
Even though I don’t use Opus 4.6 much because of the price, the feeling it gives me is that of a humanities grad — a humanities grad who can reliably read/edit large amounts of text, design products, and know what I’m trying to do.
GPT 5.4 is just a backend engineer. It can really write code, though its taste isn’t good, its product design ability isn’t great either, and it’s not great with words — can’t express things clearly.
Gemini 3.1 has something wrong with its brain this generation, some impairment in its working memory. It gets into character easily, is very edgy/chuunibyou, is very good at teaching people, but most importantly, it’s very good at flattery () If you run into something unpleasant in life, Gemini can, with sound reasoning and evidence, praise you until you’re floating on air (