In OpenFL, we include multiple renderers – OpenGL (including WebGL and GLES), Cairo, HTML5 canvas, and HTML5 DOM. We default to OpenGL on all targets and fallback to Cairo or canvas (depending on whether it is native or HTML5) when hardware is not available, or when using bitmapData.draw
OpenFL is designed to provide a broad range of Flash features. I believe we have done this without compromising performance or seriously impacting the final file size (thanks to features such as dead-code elimination and minification).
We have been focused on better support for Stage3D. This (in a sense) is less flexible than the core OpenFL display object support, because it has no software fallback. Flash Player has a software fallback for Stage3D content, but I would assume that direct HTML5 canvas calls (like our canvas renderer) would be more efficient on HTML5 than trying to simulate GL 
The goal for supporting Starling is to provide a broad set of choices for developers. Particularly, porting an existing title that already uses Starling is much simpler if there is a solid Haxe version of Starling.
The addition of Tilemap fills an important (missing) role in the core Flash API – the need for a hardware quad batch. I hope that we can continue to improve that API to feel as if it were a core API, and to continue to grow more useful.
I hope this answers some questions.
Away3D 5 was also released, I’ll be writing more in a blog post soon 