Before migrating legacy code with AI, capture what it actually does
Added to DistroCool on 2026-10-20 · Discovery date, not a release announcement
OpenAI's October 19 research reports better coding-model results in less-common languages such as COBOL and Fortran. The practical lesson: make an AI-assisted migration prove it preserves existing behavior, rather than trusting a plausible rewrite. The published scores are OpenAI's evaluations, not independent production tests.
Why it matters
Small differences in fixed-point rounding, missing records or output formatting can change a business result even when the new program runs. OpenAI released its MultiLangBench setup for inspection, but its separate migration results use hidden tests. This is a review-and-test workflow for a small non-sensitive routine, not a green light to move a critical system. The latest models are available through the API; usage has a cost, and no price for this specific experiment is promised here.
Try this next.
Copy into your coding assistant, replacing the brackets: 'Analyze this small [language] routine using only the non-sensitive code and examples I provide. Before proposing any rewrite, list its inputs, outputs, fixed-point/rounding rules, missing-record behavior and external dependencies. Write characterization tests for normal, boundary and failure cases; tell me which behavior is still unknown. Propose a minimal migration only after those tests exist. Compare old and new outputs on the same safe examples, report every difference, and do not edit or run production code.' Review the tests yourself before trusting the result.
Source: OpenAI research announcement and linked MultiLangBench repository (19 October 2026) ↗