I made my own encoder and experimented with variable block size (with dynamic programming) plus a massive amount of brute-force search for LPC parameters. But the amount of compression gain I could achieve over libflac was small and not worth the hours of encoding time. https://www.nayuki.io/page/benchmark-of-nayukis-flac-encoder

Sometimes encoding time doesn't matter. Archiving is done once in a lifetime. I encode everything with "--lax -8Vepl32" which takes ages but since it runs as a low priority background task I don't care. Also, you can use heuristics to choose the right blocksize instead of brute-forcing it. If the algorithm is clever enough you might achieve slightly better results at almost zero costs.

Interesting. Do you have numbers to show that variable block size can be done with "almost zero cost" as you said?

I think "If the algorithm is clever enough" is a big weasel phrase just like https://wiki.c2.com/?SufficientlySmartCompiler . Having studied and written many algorithms, I have to conservatively assume that a clever algorithm doesn't exist unless it is proven by construction.