COBOL migration accuracy jumps from 5% to 87% in one model generation

OpenAI’s GPT-6.1 Sol migrated COBOL programs to modern languages correctly 86.6% of the time at low reasoning effort. Its predecessor, GPT-5.6 Sol, managed 5.4% at the same setting.

The source is OpenAI’s own write-up of two new evaluations for low-resource programming languages, published on 19 October 2026.

The test is strict. A migration counts only if hidden tests match the original’s output byte for byte, so code that runs cleanly but gives different answers still fails.

The classic trap is arithmetic. COBOL fixed-point maths truncates excess decimal digits unless told to round, so a tidy rewrite that rounds can quietly change financial results.

The improvement came with a lower bill. The 86.6% run cost 61% of what the older model’s attempt did.

At higher reasoning effort the score reaches 97.9%, which OpenAI describes as close to saturating the evaluation.

The stakes are real. OpenAI cites an IBM figure that COBOL supports more than 40% of online banking systems and 95% of ATM transactions, yet it still says production use needs expert review against business rules.

Source: OpenAI, Advancing coding in low-resource programming languages