OpenAI says GPT-6.1 Sol preserved COBOL program behavior in 87% of low-effort tests and 98% at higher reasoning effort.
OpenAI’s latest reported tests suggest its models are improving at translating COBOL code into newer languages, but software experts interviewed by The Verge say the results do not establish that complex legacy systems can be safely replaced. The Verge published its report on October 26, 2026, following an OpenAI blog post dated October 19.
OpenAI tested migration into C++, Java, Kotlin, Python and Rust, with the stated goal of preserving what the original software does. In the company’s evaluation, GPT-6.1 Sol reached 87 percent behavior-preservation accuracy at low reasoning effort and 98 percent at higher effort. The report also gives low-effort results of around 86 percent for GPT-6 Astra, 21 percent for GPT-6 Sol and 5 percent for GPT-5.6 Sol. OpenAI says a migration counts as successful only if all tests pass.
Those figures come from OpenAI’s own evaluations, not a demonstration of a wholesale production-system replacement. The academics cited by The Verge noted that real systems can combine enormous codebases, multiple programming languages, varied data formats, integrations and institutional requirements. They questioned whether the benchmark reflects the work of dividing such systems into components that can be replaced safely.
OpenAI spokesperson Oscar Haines told The Verge that legacy code remains an unsolved problem. The company is gathering feedback from developers and enterprises and exploring better testing and improved model understanding of programming languages. The report also notes that Anthropic has presented Claude Code as a tool for COBOL modernization, but had not published migration performance benchmarks in the material discussed.
For organizations that depend on older systems, preserving behavior is critical: even small differences can be unacceptable. Experts emphasized thorough testing and human review; formal verification was also discussed as a possible, demanding way to check translated code. The reported benchmark results therefore indicate progress on code translation, while leaving system-level modernization and safe deployment as separate challenges.