OpenAI says GPT-6.1 Sol hits 97.9% on its COBOL migration test

Dataconomy reported on October 20 that OpenAI's GPT-6.1 Sol scores 97.9% on the company's COBOL migration test.

The figure comes from an OpenAI technical post published the day before, Advancing coding in low-resource programming languages, which examines how its models handle COBOL, Fortran and Visual Basic — languages with far less public code and documentation to learn from than Python or JavaScript.

OpenAI built two evaluations for the work. The COBOL migration test asks a model to rewrite existing COBOL programs into C++, Java, Kotlin, Python or Rust while preserving behaviour; MultiLangBench asks it to build a SQLite-compatible database engine from scratch in each of 20 languages. One measures whether a model can keep an old system doing exactly what it did, the other whether it can build a new one.

On migration, GPT-6.1 Sol reaches 86.6% accuracy at low reasoning effort, against 5.4% for GPT-5.6 Sol, at roughly 39% lower cost. Raising reasoning effort pushes accuracy to 97.9%, which OpenAI describes as nearly saturating the evaluation.

The headline number travels because the task it measures has long resisted automation. COBOL still runs large parts of core banking, card processing and ATM settlement, often in codebases whose documentation is incomplete and whose original authors have retired, leaving the business rules embedded in the source.

That is also why the grading is unforgiving: migrations are scored on hidden tests that require byte-for-byte parity on observable outputs, and a single failing case counts as a failure. OpenAI cites a characteristic trap in which COBOL's fixed-point arithmetic silently discards excess decimals, so a rewrite that rounds instead will still run while quietly changing financial results.

The jump from 5.4% to 97.9% is largely bought with inference-time compute rather than a new model generation, which raises the question enterprises actually care about: cost per migration rather than peak accuracy. The low-effort tier OpenAI reports is aimed squarely at that.

OpenAI has also published the MultiLangBench setup and per-language results on GitHub, so developers can inspect what is being tested and where their own language still falls short. For the legacy-modernisation market, that turns a vendor claim into a reproducible public benchmark.

What to watch: whether cost at high reasoning effort falls into enterprise budgets, whether the same accuracy survives real corporate codebases rather than constructed tests, and whether rival labs publish comparable COBOL migration benchmarks of their own.