GPT-6.1 Sol Takes a Big Leap on Legacy Code

Trillions of lines of COBOL still run the world's banks, insurers and government payrolls, much of it written by people who have long since retired. Migrating that code to a modern language is slow, expensive and risky, which is exactly why OpenAI chose it as a yardstick. On 19 October the company said its newest model, GPT-6.1 Sol, had gone from barely functional to nearly perfect on an internal test of that very task.

The numbers are hard to ignore. At a low reasoning setting GPT-6.1 Sol scored 86.6 percent, up from 5.4 percent for the previous GPT-5.6 Sol at the same setting, as BleepingComputer reported. Turn the reasoning effort up and the score reaches 97.9 percent, close to saturating the benchmark.

What the test measures

The eval asks a model to translate COBOL programs into C++, Java, Kotlin, Python or Rust. A translation counts only if the new program produces byte-for-byte identical output and passes every hidden test. That is a demanding bar, and it rewards getting the behaviour exactly right rather than merely close.

It is worth being clear about what this is and is not. The model itself shipped on 29 September, so this is OpenAI disclosing new evidence about an existing model, not a launch. The tests are OpenAI's own and have not been reproduced by an outside party. And the benchmark covers self-contained programs, not the tangle of a real mainframe system with its decades of patches and undocumented quirks. OpenAI's own material says production migrations still need expert review and testing.

The bottleneck moves

Take the caveats seriously and the result still matters. If a model can reliably convert isolated COBOL routines, the hard part of a modernisation project stops being the translation and becomes everything around it: working out what the old system was meant to do, untangling its dependencies, and trusting the output enough to put it into production. OpenAI also released a separate suite, MultiLangBench, on GitHub, where GPT-6 Astra built a SQLite-compatible database engine across 20 languages with an average pass rate of 87.4 percent, according to Dataconomy. Both results point the same way. The code is getting easier. The judgement around it is not.