My own experience is that for coding tasks which are probably very well represented in the training data set, such as generating a react page with some functional buttons, GPT-4 performs perfectly. When coming to more specialized tasks (with probably fewer samples in the data set for just that task), such as creating integration between systems with concurrency handling, it still performs surprisingly good but only “first draft” quality and the generated code contains bugs, misses important aspects and the code does not compile usually on the first try - even when using popular programming languages, popular libraries and effort is put in to describing the problem&expected results.
Based on this, an excellent tool for developers, but not ready to replace them even though it is surprisingly good. For scaffolding tasks that junior programmers do - it could possibly cut down the need for them (in additio to providing valuable assistance to non-scaffolding standard tasks).