Which model is best for legacy code?

Factory's Legacy-Bench, now included in Fireworks' Specialized Intelligence Index, asks frontier models to debug and migrate code in COBOL, Java 7, BASIC, C89, Fortran and Assembly. Scores span from 60% down to 23% across the models tested, a much wider spread than modern-language coding benchmarks tend to show. The finding that general coding strength does not carry over cleanly to older languages matters for enterprises weighing AI-assisted migrations. The most useful detail is a failure case: one model silently miscalculated a payroll deduction while still passing most of the test suite. That is the risk profile of legacy migration in one example, since the code often encodes business rules that tests only partly capture. Teams using models on this kind of work should add property checks or parallel runs against the original system for calculations that touch money. Separately, Factory reported its model router now cuts production inference costs by 63%, up from 42% in June.