OpenAI's October 19 post covers two evaluations. MultiLangBench asks a model to build a SQLite-compatible database engine in each of 20 languages, scored with tests adapted from SQLite's own suite. The second asks a model to migrate COBOL programs to C++, Java, Kotlin, Python or Rust, and a migration only counts if hidden tests show byte-for-byte matching output.
On MultiLangBench, GPT-6 Astra averaged an 87.4% test pass rate at maximum reasoning effort, against 26.8% for GPT-5.6 Sol. COBOL was still its weakest language, at 72.1%. GPT-6 Luna, the cheaper model, averaged 40.3% at an estimated $0.77 per sample, versus $3.48 for GPT-5.6 Sol.
On the COBOL migration eval, GPT-6.1 Sol scored 86.6% at low reasoning effort (GPT-5.6 Sol got 5.4%) at 61% of the API cost, and 97.9% at higher effort. OpenAI gives one example of what goes wrong in these rewrites: COBOL's fixed-point math drops extra decimal digits unless rounding is requested, so a rewrite that rounds can run without errors and still produce different financial results. It says production migrations still need expert review. All the numbers are OpenAI's own, but the MultiLangBench setup and per-language results are public.