Comment by samsmallmind

100% agree, have found through tests a big discrepancy between models optimised for SWE benchmarks and usage in prod - there is a somewhat subjective element to it as well in terms of output that can't be quantified effectively (less so with decision models naturally)