After testing out Claude Opus extensively via Perplexity AI, I can say it is significantly better than GPT4 at coding.
One of the best benchmark test to do is request the model to design a snake game in Python, I like to request at least 10 enhancements that the model will decide on and include in the game. Claude Opus got a fully working Snake game with 9 of 10 enhancements working in a single prompt request (it added obstacles and walls in the game, the only thing was the high score wasn't visible being black on black). GPT 4 did have the game run, for some reason chose WSAD as controls over arrow keys, but the snake moved like a turtle with no reset after a game over. Overall both worked but Claude Opus implementation was definitely much better.
I proceeded to have Claude Opus build an app from scratch with a GUI and dropbox interface capturing current weather for any large city requested. It recommended a free API weather option and build the app designed around the API for temperature, windchill / humidex, humidity, conditions, atmospheric pressure. After registering for an API I was able to successfully get a fully working weather app. I haven't tested GPT4 but from experience it will have some issues with this requiring additional steps.
Claude Opus definitely does have its censor barrier, but not nearly as bad as Claude 2 was reported to have, and way better than Gemini Pro which is a preaching and lecturing disaster.
A great thing about Perplexity as a tool is that you can switch models in the middle of conversions (say you hit a censor barrier with Opus, just switch to Mistral Large), add files, images, create images, do web searches, all in a very convenient and efficient format. When building the weather app, it read through the API guide, and proceeded to fix its code using the guide and the app worked. The app literally took 20 min to build and that includes registering and subscribing for the API. After using Perplexity, outside of running my GPTs I custom built, I am using it as my go to tool.
|