OpenAI's benchmarks show GPT-6 Astra lifting COBOL from a 10.9% to a 72.1% pass rate and GPT-6.1 Sol hitting 86.6% on COBOL-to-Java migration, but anyone who has worked a COBOL migration knows the hard part isn't syntax.
It's arithmetic, since a rewrite that rounds fixed-point values instead of dropping the extra digits runs cleanly and still produces different financial results, and the benchmark only counts because someone already wrote the hidden tests a real mainframe estate does not have.
See why the test harness, not the model, decides how fast teams can cut over: buff.ly/ZfBztqt
#AI #OpenAI #COBOL #Mainframe #LegacyModernization #DevOps
It's arithmetic, since a rewrite that rounds fixed-point values instead of dropping the extra digits runs cleanly and still produces different financial results, and the benchmark only counts because someone already wrote the hidden tests a real mainframe estate does not have.
See why the test harness, not the model, decides how fast teams can cut over: buff.ly/ZfBztqt
#AI #OpenAI #COBOL #Mainframe #LegacyModernization #DevOps
Oct 21, 2026 · 7:30 PM UTC
1
0
1
372
@devopsdotcom Test the arithmetic first. COMP-3 fields map to BigDecimal with a fixed scale, and COBOL truncates unless ROUNDED is set, so the Java side needs RoundingMode.DOWN, not HALF_UP. Replaying a month of real mainframe inputs and diffing the outputs beats a benchmark.
0
0
0
1