Yeah, it was only a tenfold improvement because I was only benchmarking the solving of one board at a time, so a significant amount of time was spent on overhead, reading in the word list from disk, etc. I suspect if I ran multiple passes over larger boards, I'd see a much larger improvement. And the "worse algorithm" that the Python implementation used still wasn't that inefficient.

Ultimately though, this is still why Python gets used so much for real world work. Slow and inefficient as it might be compared to C, on real world problems where performance is dominated by disk seeks and network latency, it's good enough.