OpenAI announced on 20 October 2026, according to Engadget, that its AI models are now better equipped to handle software building and migration in aging programming languages. This development targets legacy systems like COBOL, Fortran, and Visual Basic that still run significant portions of global financial infrastructure.
Key takeaways
- Engadget reports that 40 percent of the world's banks still rely on aging COBOL systems.
- OpenAI tested its models using a new evaluation suite named MultiLangBench.
- GPT-6 Astra achieved a test pass rate of 87.4 percent across 20 languages.
- GPT-6.1 Sol reached up to 97.9 percent accuracy when migrating COBOL programs to modern languages.
How does OpenAI COBOL migration work?
As Engadget explains, banking institutions face steep costs and time commitments when trying to move away from legacy code. Because the programmers who know these languages are aging out of the workforce, financial firms need to update their systems soon. To track progress, OpenAI built an evaluation suite called MultiLangBench. This benchmark tests models by asking them to create SQLite-compatible database engines in specific programming languages.
During tests, GPT-6 Astra hit an average pass rate of 87.4 percent across 20 languages. That score marks a major jump compared to GPT-5.6 Sol, which only reached 26.8 percent. Even though COBOL remains the hardest language to crack, Astra still achieved a 72 percent pass rate on it.
Testing software migration tools
OpenAI also tested how well its models can translate COBOL programs into modern languages like C++, Java, Kotlin, Python, or Rust. The main challenge is keeping the exact program behaviors intact, such as how excess decimal digits are rounded or how missing records are handled.
In these specific migration tests, GPT-6.1 Sol reached an accuracy of 97.9 percent when running with high reasoning effort. When switched to low reasoning effort, the same model dropped to 86.6 percent accuracy.
What it means for financial institutions
Engadget notes that OpenAI acknowledges that any software update only works if all tests pass. Financial software demands extreme precision, which means teams cannot rely entirely on automation yet. OpenAI stated that behavior-preserving capabilities help teams maintain older systems, but putting that code into production software still requires expert review and rigorous testing against business rules and dependencies.
Frequently asked questions
Engadget reports that 40 percent of the world's banks still rely on COBOL systems.
MultiLangBench is an AI evaluation suite created by OpenAI that asks models to create SQLite-compatible database engines in a specified programming language.
Engadget states that GPT-6 Astra had an average test pass rate of 87.4 percent across 20 languages.
GPT-6.1 Sol achieved the highest accuracy for migrating COBOL programs, reaching 97.9 percent accuracy with high reasoning effort.
No, OpenAI notes that applying migrated code to production software still requires expert review and testing against business rules and dependencies.
Based on reporting by Engadget. Read the original article
