OpenAI announced that its smart models have become more efficient in building and updating software written in legacy programming languages facing a declining number of proficient programmers, such as COBOL, Fortran, and Visual Basic. Company data indicates that about 40% of banks worldwide still rely on systems based on COBOL, and although transitioning to modern languages costs millions of dollars and takes a long time, the retirement of specialized programmers forces these institutions to undertake a technical migration over the next few years.
The company based these capability developments on feedback and input from developers and enterprise customers, as programming models previously faced real challenges when processing legacy programming languages. To evaluate performance, OpenAI developed a benchmark suite known as MultiLangBench, which requires models to build SQLite-compatible database engines in a specific programming language.
Test results showed significant superiority, with the GPT-6 Astra model recording a success rate of 87.4 percent across 20 languages, vastly outperforming the GPT-5.6 Sol model, which recorded 26.8 percent. Despite this improvement, COBOL remained the toughest challenge, with Astra's success rate reaching about 72 percent.
Why This News Matters
Legacy banking systems represent a major obstacle to technological development due to their association with critical financial data and complex business rules that require extreme precision. The significance of the new development lies in the models' ability to preserve original program behaviors during the conversion process to languages such as C++, Java, Kotlin, Python, or Rust.
In conversion evaluation operations that strictly preserve program behavior—such as the method of rounding complex decimal numbers and handling missing records—the GPT-6.1 Sol model achieved the highest accuracy of 97.9 percent when using a high level of reasoning, while accuracy reached 86.6 percent at low reasoning levels.
Limitations and Practical Requirements
OpenAI emphasizes that the success of any software migration process requires passing all accuracy tests, given the sensitivity of banking tasks. The company notes that these capabilities serve as a valuable assistance tool for maintenance and modernization teams, but deploying software in actual production environments still requires specialized human review and rigorous testing conforming to business rules and system-related considerations.