OpenAI's GPT-6.1 Sol hits 97.9% on COBOL migration, up from 5.4% — MultiLangBench open-sourced

#What happened

OpenAI published a research post on October 19, 2026, "Advancing Coding in Low-Resource Programming Languages," reporting two new evaluations for legacy code work and releasing one as an open benchmark. MultiLangBench asks a model to write a SQLite-compatible database engine in one of 20 programming languages, graded against converted SQLite test cases, and the setup, suite, and per-language results are published on GitHub. The COBOL migration test asks models to translate COBOL programs into C++, Java, Kotlin, Python, or Rust, counting a migration only when every hidden test passes byte-for-byte.

#The numbers

COBOL migration: GPT-6.1 Sol scored 86.6% accuracy at low effort and 97.9% at higher effort, versus 5.4% for the older GPT-5.6 Sol baseline. MultiLangBench (average across 20 languages): GPT-6 Astra reached an 87.4% pass rate and GPT-6 Luna 40.3%, versus 26.8% for GPT-5.6 Sol. Astra scored above 83% in 19 of 20 languages, with COBOL still the hardest; Fortran jumped from 21.4% to 83.5% and Visual Basic from 27.5% to 89.4% with Astra. Luna's results came at about $0.77 per sample versus $3.48, roughly 78% cheaper, and GPT-6.1 Sol's low-effort COBOL migration cost about 61% of the older model's API cost.

#Why it matters

COBOL still runs much of the world's banking and payments infrastructure, and legacy-code teams are the bottleneck these evals target. OpenAI itself cautions that these are its own benchmarks, not independent tests, and that real production migrations still need expert review. The open MultiLangBench release lets outside developers check the claims and run rival models through the same tasks.

Verified October 19, 2026. Sources: AI Parabellum (aiparabellum.com/openai-cobol-legacy-code-benchmark) and Blockchain.News (blockchain.news/news/openai-low-resource-programming-languages-evaluations), both reporting on OpenAI's October 19, 2026 research post "Advancing Coding in Low-Resource Programming Languages" and its MultiLangBench GitHub repository.