OpenAI says GPT-6.1 Sol COBOL migration is now close to solved on its own test: the model converted COBOL programs to modern languages with 97.9% accuracy at higher reasoning effort. The older GPT-5.6 Sol managed just 5.4% at low effort. The catch is that OpenAI built and ran the benchmark itself.
This matters if you work anywhere near banks, insurers, or government systems, where old COBOL code still runs quietly in the background.
Here are the quick facts, so you can skim first.
- Announced on October 20, 2026, by OpenAI.
- Task: rewrite COBOL in C++, Java, Kotlin, Python, or Rust without changing behavior.
- GPT-6.1 Sol scored 86.6% at low effort and 97.9% at higher effort.
- The low-effort run cost about 61% of what GPT-5.6 Sol cost.
- All numbers come from OpenAI’s own evaluations, not an outside lab.
The short version: a big jump with real caveats. For anyone planning legacy code migration, the details of the test matter as much as the score, so we will walk through them next.
How the GPT-6.1 Sol COBOL test works, and why it is hard
OpenAI’s benchmark asks a model to translate a COBOL program into another language. Hidden tests then check that the output matches the original byte for byte. One wrong digit and that program counts as a failure.
That strictness is the point. COBOL uses fixed-point arithmetic that quietly drops extra decimal digits unless the code asks for rounding. A rewrite that rounds “helpfully” can run without errors and still produce different money totals.
So a high score here is harder to fake than on a typical coding benchmark, where a program that merely runs can pass. Converting COBOL to Python or Rust also means matching old data formats, not just logic.
The numbers, side by side
| Test | GPT-5.6 Sol | GPT-6.1 Sol |
|---|---|---|
| COBOL migration, low reasoning effort | 5.4% | 86.6% |
| COBOL migration, higher reasoning effort | Not reported | 97.9% |
| Cost of the 86.6% run vs. GPT-5.6 Sol | 100% (baseline) | About 61% |
The jump from 5.4% to 86.6% is large. OpenAI itself describes the 97.9% result as close to saturation, meaning the test has little room left to measure improvement. TechBriefly reports the low-effort run cost 61% of GPT-5.6 Sol’s API bill. That is the number most teams will care about, since few want to pay for maximum reasoning on every file.
Why old languages are suddenly a big deal
OpenAI calls COBOL, Fortran, and Visual Basic “low-resource” languages. There is far less public code for them than for Python or JavaScript, so models usually learn them poorly.
The stakes are high anyway. The Decoder cites an IBM estimate that COBOL runs over 40% of online banking systems and 95% of ATM transactions. Meanwhile, the pool of engineers who can maintain it keeps shrinking.
OpenAI also released a second test, MultiLangBench, on GitHub. It asks models to build a SQLite-compatible database from scratch in 20 languages. GPT-6 Astra at maximum effort averaged 87.4%. TechBriefly adds that COBOL was the hardest language of the 20, which fits everything above about its quirks.
What this does not prove
A benchmark is not a bank. The tests check program output, but real mainframe systems depend on job schedulers, database links, and business rules no test file captures.
OpenAI says as much. Production use “still requires expert review and testing against the system’s business rules and dependencies,” according to the reports. No independent group has reproduced the scores yet, and the coverage does not give pricing or rollout details for the newer models.
If you want background on the model family, our guide to GPT-6 in ChatGPT and who gets Sol and Luna covers the tiers. For another case of AI tackling old, awkward code, see how Claude Code wrote a Linux driver for an unsupported scanner.
What developers and IT teams can do now
You do not need to wait for a perfect score to start testing. Here is a sensible way to try it on your own code.
- Pick one small, well-understood COBOL module, not your core ledger.
- Keep the original running and compare outputs on real historical data.
- Check rounding, date handling, and number formats first, since those fail quietly.
Running both versions side by side is what turns an AI rewrite from a gamble into an experiment.
Frequently Asked Questions
Can GPT-6.1 Sol replace COBOL programmers?
No. It can speed up translation, but people still need to review the output and test it against business rules. The skill of understanding what the old system should do becomes more valuable, not less.
Is the 97.9% COBOL score independently verified?
Not yet. OpenAI designed and ran the evaluation, so treat the figure as a vendor claim until outside teams repeat it.
Which languages can it convert COBOL into?
The test covered C++, Java, Kotlin, Python, and Rust.
Why is COBOL still used?
It runs huge, reliable financial systems that are costly and risky to rewrite. Banks keep it because it works, not because it is fashionable.
Our take
This is a genuine step forward for legacy migration, and the byte-for-byte scoring makes it more convincing than most coding claims. But I would still treat it as a strong assistant, not an autopilot. If your team owns old COBOL, run a small pilot this quarter and measure the results yourself.
Leave a Reply