#GPT 4.1 vs Claud 3.7 Sonnet - How have you found it?

1 messages · Page 1 of 1 (latest)

mighty sedge
#

So while we wait.. how has everyone found 4.1 vs Claude?

tender eagle
#

it depends on the usage.... for complex tasks claude 3.7 thinking is better for quick raw drafts GPT is great

mighty sedge
#

I've found GPT did well planning out a project and posing followup questions before scaffolding

spare sun
#

I find Claude 3.7 often works like a younger over achiever that tries to do too much, over complicates it, loses track of the work and then as the project gets bigger fails in an unrecoverable way.

GPT 4.1 seems to move at a more medium pace, always asking me if it wants me to complete something or telling me to do the things and click the buttons, and again as projects get bigger, it in a more slow way ends up dissapointing me.

However I am hopeful as AI is so much better than it was last year at this time, and it was laughably horrible 2 years ago, so I am neutrally satisfied with my dissapointing projects.

mighty sedge
#

That's a really interesting perspective on Claude because I've found the same thing sometimes.

spare sun
#

I am currently for whatever crazy reason burning credits with gemini pro 2.5 and still just spinning my wheels, I have used rules, documents, index.md map.md files, mcp servers, and memories, the ai doesnt ususally automatically use mcp servers when it needs to, or trigger memories automatically unless mentioned, and has trouble deploying stacks in docker for complex workflows like all the other AI models.

mighty sedge
#

I haven't used GP2.5 but I've heard it does quite well. Are you saying its not actually as good as people are saying?

fast flare
#

if need something creative but not complicated - gpt 4.1, if coding, parsing, etc. - claude 3.7 (obviously, thinking)

spare sun
#

its fine for individual projects or somewhat simple workflows. but for full stacks and ai integration into those apps and full feature developement, they all kinda suck

#

I wish Windsurf has a great custom rules system like Cursor, Awesome modes and model switching like Roo Code, Great context management like Augment Code and great MCP tool use like Cline. If I could get all of that together in 1 IDE with 1 app AI wouldn't suck at all. I wish I knew someone at Windsurf that I could say this to, because it could change the world. ;p

blazing ridge
#

I replaced 3.7 and gemini and use exclusively gpt 4.1. With proper instructions it works perfectly for me. Here is mine for behavior... (I have other for specific tech stack and coding conventions). It's a markdown format:

General Behavior

Always use the following behavioral guidelines when answering user requests.

Role

  • You are an senior level coding agent with an expertise in full stack web development technologies.

  • Your primary use is to help efficiently design, code, debug, and optimize web applications.

Behavior

  • Your thinking should be thorough and so it's fine if it's very long.

  • You MUST plan extensively before each function call, and reflect extensively on the outcomes of the previous function calls. DO NOT do this entire process by making function calls only, as this can impair your ability to solve the problem and think insightfully.

  • Please keep going until the user’s query is completely resolved, before ending your turn and yielding back to the user. Only terminate your turn when you are sure that the task or problem is solved.

  • If you are unsure about the file content or the structure of the codebase relevant to the user’s request, use your tools to examine the files and gather the necessary information, or ask the user for additional details: do NOT guess or make up an answer.

Code specific behavior

1. Respect the Code Context

Analyze the surrounding code from the file you are working in (if available), and ensure that your response integrates seamlessly with it. Consider dependencies, existing conventions, and architectural patterns before writing or modifying code.

2. Follow Modern Best Practices

Ensure that all code you generate adheres to up-to-date best practices for the given language, framework, and task. This includes naming conventions, structure, safety, performance, and maintainability.