I made 4000 agent calls in Cursor last month. Each model has a personality

The lazy architect (OpenAI’s o3). o3 is incredibly lazy at writing code, but very good at planning. Will happily read tens of files and do deep analysis, but often struggles in scenarios where it needs to edit more than one file.

The over-eager child (Claude Sonnet 3.7 Thinking). Claude Sonnet is eager to just get going, man! It’s not the most careful, and in longer strings of tool calls, may start editing something completely unrelated to what you asked it to.

Pretty balanced?(Gemini 2.5 Pro). Gemini 2.5 is a little more intelligent, and significantly faster and more reserved than Sonnet 3.7. Usually the best choice for writing code in multiple files.

I’ve found o4-mini to be incredibly slow and fairly mediocre, and GPT 4.1 useful in very situational areas. My tips:

- Use o3 to plan and/or write code in one or max two file only. If you do more, it may openly revolt and just refuse to write any longer.

- Always make sure Sonnet 3.7 is following a tightly scoped plan on a relatively small section of the product, and supervise it. If you have an easy change to make in many areas of your codebase, for example, letting Sonnet run, still supervised, is a perfect use of the model’s persona

Generally what I do:

- Medium complexity: editing one file: o3. Editing multiple files: plan with o3, write with gemini-2.5

- Simple complexity: Editing many files, very simple: plan with o3 if needed, write with claude-3.7. Editing many files, simple, needs formulaic approach: write a detailed prompt into GPT 4.1

- High complexity: plan with o3, separate into multiple chunks, write small chunks at a time with gemini-2.5 and be very careful with each section. If I'm super lazy sometimes I just YOLO all of the sections and then fix all the bugs at the end but this probably leads to code issues later down the line.

Would love to hear other people are using the different models!