It's not about performance but about convenience. Having a way to plot single pixels with the CPU without a lot of boilerplate code to setup a 3D API context, creating textures, vertex buffers, shaders etc etc is a good thing to have in many situations.
That this is not a good way for doing more complex rendering tasks seems quite obvious (e.g. in the long run this cannot be the only API for covering all sorts of rendering, but it happens to be a good sweet spot for getting anything on screen at all, and not having to deal with an overly complex rendering API just to get some 2D image data rendered).
If convenience is what you want, that can be provided by a 2D canvas. This could even support a "soft framebuffer", i.e, a pixel-perfect canvas taking up the full screen, or some well-defined window inside it. But that would not be a low-level rendering API, of course.
To reinforce on this: is pixel actually a well defined usable unit? as far as I understood many modern monitor often use a definition of pixel quite different from the one of a couple decades ago.
That this is not a good way for doing more complex rendering tasks seems quite obvious (e.g. in the long run this cannot be the only API for covering all sorts of rendering, but it happens to be a good sweet spot for getting anything on screen at all, and not having to deal with an overly complex rendering API just to get some 2D image data rendered).