If you really have a matrix mulitplication algorithm that multiplies nxn matrices in less than O(n^2.3729), you have made a mathematical breakthrough and should publish.

I second "publish".

If you've invented something worth noting, you'll get infinitely better residual value, visibility, respect and goodness coming your way for years to come vs. one time deal of selling your work to a private buyer for a one time fee in a sell-and-forget type of deal.

What do you mean by "with better accuracy"? How can one matrix multiplication algorithm be more accurate than another?

Strassen makes your rounding errors bigger the more levels of recursion you do. We're talking ~7 bits of error due to rounding (in a paper I read).

Real computers are finite precision. There are very real issues with numerical stability where you are running along fine for many many cases, and then there is a catastrophic cancellation that wipes out all of your accuracy. There is a huge difference between writing a paper on bounds of matrix multiplication, and delivering performance in a finite precision world.