OpenAI says GPT-6 Astra and GPT-6.1 Sol improved on its new coding evaluations for building SQLite-compatible software and migrating COBOL programs.
OpenAI has published results from two evaluations designed to measure model performance on software tasks in programming languages with comparatively limited training material, including COBOL, Fortran and Visual Basic. The company’s October 19, 2026 report covers building a SQLite-compatible database engine across 20 languages and migrating COBOL programs while preserving their behavior.
On MultiLangBench, OpenAI reports that GPT-6 Astra achieved an average test pass rate of 87.4% across the 20 languages at maximum reasoning effort. That compares with 26.8% for GPT-5.6 Sol. Astra scored above 83% in 19 languages; COBOL was the most difficult, with a 72.1% pass rate. OpenAI also reported improvements in OCaml, Fortran and Visual Basic. GPT-6 Luna averaged 40.3%, with an estimated API cost of $0.77 per sample, compared with $3.48 for GPT-5.6 Sol—about 78% lower, according to the company.
The second evaluation tests translations from COBOL into C++, Java, Kotlin, Python or Rust. OpenAI says GPT-6.1 Sol reached 86.6% accuracy at low reasoning effort, versus 5.4% for GPT-5.6 Sol at the same setting, while costing 39% less. At higher reasoning effort, GPT-6.1 Sol reached 97.9% accuracy. The evaluation uses hidden tests and requires byte-for-byte parity on observable outputs for a migration to count as successful.
These tests target two distinct maintenance challenges: writing new software in less commonly supported languages and changing legacy code without altering its behavior. That distinction matters for older systems where business rules may be embedded in code or where details such as fixed-point arithmetic affect financial results. OpenAI cautions that production migrations still require expert review and testing against the system’s business rules and dependencies.
OpenAI says it is releasing the MultiLangBench setup and test suite, along with results by language. The published scores are company-reported evaluation results, rather than a guarantee that a model will preserve behavior in any particular production codebase.