COBOL and Fortran stop being blind spots: new models lift SQLite rebuild pass rates from 26.8% to 87.4% across 20 languages

Coding in low-resource languages now has a quantified comparison: COBOL, Fortran and Visual Basic have less code and documentation for models to learn from than Python and JavaScript, yet COBOL supports more than 40% of online banking systems, 80% of in-person credit card transactions and 95% of ATM transactions. Developers and enterprise customers had reported that coding models often struggle with these languages; maintaining and modernizing such software is hard because documentation may be incomplete, original developers may have left, and business rules may exist only in code. On MultiLangBench, GPT-6 Astra averages an 87.4% test pass rate across 20 languages versus 26.8% for GPT-5.6 Sol, with gains in every language tested, including COBOL from 10.9% to 72.1%, OCaml from 25.6% to 85.8%, Fortran from 21.4% to 83.5% and Visual Basic from 27.5% to 89.4%.

A lower-cost option also beats GPT-5.6 Sol in every language tested, so the gains are not confined to the maximum-compute setting. GPT-6 Luna reaches a 40.3% average test pass rate at maximum reasoning effort, below Astra but above GPT-5.6 Sol's 26.8%. GPT-6 Luna costs an estimated $0.77 per sample on average versus $3.48 for GPT-5.6 Sol, about 78% lower; its per-language results include COBOL from 10.9% to 17.1%, Fortran from 21.4% to 31.4% and Visual Basic from 27.5% to 41.6%.

The second evaluation tests whether behavior survives migration, not merely whether the code runs. COBOL programs often process financial transactions and business records, where rounding rules and missing-record handling reflect business requirements; COBOL fixed-point arithmetic discards excess decimal digits unless rounding is requested, so a rewrite that rounds can run without errors yet produce different financial results. The migration evaluation requires byte-for-byte parity on observable outputs in hidden tests, and a migration succeeds only if all tests pass; GPT-6.1 Sol reaches 86.6% at low reasoning effort versus 5.4% for GPT-5.6 Sol at the same setting, at 61% of the API cost, and accuracy reaches 97.9% at higher reasoning effort.

The results concern two in-house evaluations: MultiLangBench asks a model to build a SQLite-compatible database engine in a specified language, using database files as a shared interface and a reference SQLite implementation to check resulting contents; the COBOL migration evaluation asks for migration to C++, Java, Kotlin, Python or Rust with byte-for-byte parity on hidden tests. The intended readers are developers and enterprise teams maintaining and modernizing legacy systems. The MultiLangBench setup and test suite are released on GitHub, with results viewable by language.

Migration accuracy reaches 97.9% at higher reasoning effort, close to saturating the evaluation, leaving limited headroom. COBOL remains the hardest of the 20 languages on MultiLangBench, at 72.1% for Astra, while the other 19 are above 83%. Applying such behavior-preserving capability to production software still requires expert review and testing against the system's business rules and dependencies.