TL;DR: OpenAI published results on Oct 19, 2026 showing its newest models are much better at older, less common programming languages such as COBOL, Fortran and Visual Basic. On OpenAI’s own COBOL migration test, GPT-6.1 Sol got 86.6% of programs fully right at low reasoning effort, against 5.4% for GPT-5.6 Sol, and 97.9% at higher effort. OpenAI also released MultiLangBench, its test of whether a model can build a SQLite-compatible database engine in 20 languages. If your business still runs on legacy code, you can now check the scores for your language before you trust a model with it.

What OpenAI announced

OpenAI calls COBOL, Fortran, OCaml and Visual Basic “low-resource” languages: there’s much less public code and documentation for models to learn from than for Python or JavaScript. OpenAI says the work was guided by developers and enterprise customers who told it coding models “often struggle with these languages.”

It reported results on two of its own evaluations.

COBOL migration. Models convert COBOL programs to C++, Java, Kotlin, Python or Rust, and hidden tests check for “byte-for-byte parity on observable outputs.” A migration only counts if every test passes.

  • GPT-6.1 Sol: 86.6% at low reasoning effort, compared with GPT-5.6 Sol’s 5.4% at the same setting, at 61% of the API cost.
  • GPT-6.1 Sol at higher effort: 97.9%, which OpenAI says is close to saturating the test.

MultiLangBench. Models build a SQLite-compatible database engine from scratch in 20 languages, checked against tests adapted from SQLite’s own suite.

  • GPT-6 Astra (maximum reasoning effort): an average pass rate of 87.4%, up from 26.8% for GPT-5.6 Sol, and above 83% in 19 of the 20 languages. COBOL is still the hardest, at 72.1%.
  • GPT-6 Luna, the cheaper option: 40.3% on average, at an estimated $0.77 per sample against $3.48 for GPT-5.6 Sol, about 78% lower.

OpenAI has published the MultiLangBench setup and test suite on GitHub, so you can run it yourself.

What it means for teams with legacy code

These are OpenAI’s own tests of OpenAI’s own models, not an independent comparison, and no other vendor’s models are included. Still, the jump on COBOL migration is large, and the reason it’s hard is the kind of bug that matters to a business. OpenAI’s example: COBOL’s fixed-point math drops extra decimal digits unless rounding is requested, so a rewrite that rounds instead “can run without errors while producing different financial results.”

OpenAI is clear that the results don’t remove the need for people: applying this to production software “still requires expert review and testing against the system’s business rules and dependencies.”

What to do

  1. Check your language first. OpenAI’s post reports scores by language, and they vary a lot: GPT-6 Astra scored 72.1% in COBOL, 83.5% in Fortran, 85.8% in OCaml and 89.4% in Visual Basic. If your language isn’t listed, you can run the MultiLangBench suite yourself.
  2. Write the tests before the migration. Capture what the old program outputs today, including rounding and missing-record cases, and require the new code to match exactly. That’s the standard OpenAI’s own test uses.
  3. Start with low effort, then raise it. OpenAI’s numbers suggest low reasoning effort already does well on COBOL migration with GPT-6.1 Sol, which keeps costs down. Use higher effort on the programs that fail.
  4. Keep a person in the loop. Have someone who knows the business rules review every migrated program before it touches real money or customer records.

If you plan to run long migration sessions, our guide to Claude and Codex usage limits explains how plan limits work. For automated review on the pull requests that come out of a migration, see our coverage of Claude Code Review.

Sources & method

Sources

Method. Last checked October 20, 2026. All scores, costs and model names come from OpenAI’s post; the benchmark link comes from OpenAI’s GitHub. We haven’t run the benchmark ourselves. No affiliate links on this page. How we work.

Spotted an error? See our corrections policy and tell us.