- Who it helps
- Developers maintaining older business systems and researchers comparing coding evaluations.
- Availability
- The October 19 report links a public MultiLangBench setup. Models are available through provider APIs; the COBOL test’s hidden cases do not become your production acceptance tests.
Compare the outputs
The COBOL task checks byte-for-byte observable outputs across language migrations. The report gives 86.6% accuracy for GPT-6.1 Sol at low effort, compared with 5.4% for GPT-5.6 Sol.
Keep business rules explicit
A migration can compile while changing rounding or missing-record behavior. Review your own test cases, data formats and dependencies before using the benchmark result to justify a rewrite.
What you can do
- Use a small, representative program and documented expected outputs when assessing a migration proposal. Open reference ↗
Where the evidence stops
The benchmark and migration were not run here. Provider scores do not certify a production application.
Sources read. Public sources checked October 28, 2026. No software installation, inference or complete workflow was executed.
Sources & further reading
Web · official
Low-resource programming-language evaluations ↗
Direct evidence for the subject, its availability and the limits of its claims.
OpenAI · page checked 2026-10-28
Coverage note: Original X access was unavailable. Independent execution, comparable social measurements and complete lookback coverage were not established.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
OpenAI reports stronger results on coding languages with less training material, using SQLite-compatible programs and COBOL migration tests.