gpt-5-high is the best model for developers
Categories:
If you need to code, gpt-5-high is currently the only model that can truly boost your efficiency.
I have 10 months of experience using the Claude model and have used Gemini/DeepSeek/glm/grok sporadically. I really dislike models that don’t think. I have to admit that Claude can work for long periods and is good at using tools, but its error rate is high, leading to low usability of the results, and it often requires adjustments. However, Claude’s adjustment ability is very poor; it will repeatedly, extensively, disruptively, and superfluously wander through the codebase, crapping all over the place. After having to hard reset several hours of work multiple times, I’ve developed a deep aversion to this kind of “busy but dumb” behavior from Claude. Its working style is suitable for research, producing content that looks plausible but doesn’t hold up to scrutiny—essentially, fluff pieces. Or for operating browsers, executing tools, writing small scripts, or making minor page edits. That’s its ceiling.
I won’t discuss prompt usage here. If there are folks who are good at using Claude, by all means, continue. It’s just that Claude and I might not be compatible and can’t work together well.
Next, the only coding model I recommend is gpt-5-high, and that’s the only one. The nano, mini, medium versions of gpt-5, as well as gpt-codex, gpt-codex-high, gpt-codex-max, etc., are not on the recommendation list. They are not in the same league as gpt-5-high. Any model just shown as “gpt-5” is not gpt-5-high. With more or fewer characters, it’s not “gpt-5-high”.
The model whose behavior is most similar to gpt-5-high is OpenAI’s o3 model. If you don’t have access to gpt-5-high, using o3 a few times can also give you a feel for the model’s intelligence. I’d like to point out that VS Code GitHub Copilot used to have o3, but due to VS Code’s crappiness, it could only be used for “ask” within VS Code, and each session consumed 5 premium requests. During my long Copilot subscription, I never used o3. It was only after switching to Cursor that I discovered how high o3’s intelligence level is. VS Code GitHub Copilot has delisted o3 and claims that gpt5.1 can be a replacement. Responsibly speaking, with or without “high”, they are completely different things. I strongly recommend ditching Copilot directly. If you’re a student writing small projects with a limited budget, Copilot can still get you started.
gpt5high can be used in Cursor and in ChatGPT Plus’s codex CLI. Be careful not to select gpt-5-codex-high; it’s not the same thing.