OpenAI models show improved capabilities for migrating legacy COBOL banking systems
New benchmarks reveal OpenAI's latest models can accurately translate COBOL code to modern languages, addressing critical workforce gaps in financial institutions.
Key points
- OpenAI reports its GPT-6 Astra model achieved a 72 percent pass rate on COBOL tasks, the highest among tested languages.
- The company notes that 40 percent of global banks still rely on COBOL systems, creating urgent migration needs as experts retire.
- In migration tests, GPT-6.1 Sol reached 97.9 percent accuracy in preserving program behavior when converting COBOL to modern languages.
OpenAI announced that its latest artificial intelligence models possess significantly improved capabilities for building and migrating software written in legacy programming languages. According to Engadget, the company highlighted enhancements specifically designed to assist financial institutions and government agencies in transitioning away from aging codebases like COBOL, Fortran, and Visual Basic.
The announcement follows earlier reporting by Threat Times on October 20, 2026, which noted that OpenAI models had already demonstrated performance leaps in these legacy languages. The current update provides specific benchmark data and contextualizes the business urgency for banks that continue to depend on these older systems for core operations.
Context and Background
OpenAI stated that approximately 40 percent of the world's banks still rely on COBOL systems for their infrastructure. Migrating these environments to newer programming languages involves costs ranging from millions to hundreds of millions of dollars, alongside significant time investments. The company emphasized that the workforce of programmers skilled in COBOL is aging out, creating a pressing timeline for institutions to complete migrations within the next few years.
To measure progress, OpenAI developed an evaluation suite called MultiLangBench. This tool asks models to create SQLite-compatible database engines in specified programming languages. The company reported that GPT-6 Astra achieved an average test pass rate of 87.4 percent across 20 languages, a substantial improvement over GPT-5.6 Sol, which recorded a 26.8 percent pass rate.
Despite these gains, COBOL remains the most difficult language for the models to master. GPT-6 Astra achieved a 72 percent pass rate on COBOL tasks, which was the lowest among the 20 languages tested. OpenAI indicated that this work was guided by feedback from developers and enterprise customers who reported that existing coding models struggled with older languages.
Who Is Affected
The primary beneficiaries of these improvements are banking institutions and other organizations maintaining legacy software. For these entities, successful migration requires preserving the exact behavior of original programs, including specific handling of number rounding and missing records. OpenAI evaluated its models by having them migrate COBOL programs to C++, Java, Kotlin, Python, or Rust while maintaining functional integrity.
In these migration tests, GPT-6.1 Sol achieved the highest accuracy, reaching 97.9 percent when utilizing high reasoning effort. At low reasoning effort, the model achieved 86.6 percent accuracy. These results suggest that AI tools can assist in the complex process of modernizing critical financial infrastructure, though the company cautioned that expert review remains essential.
What should you do?
Organizations relying on legacy systems should monitor vendor announcements regarding AI-assisted migration tools. Security and development teams need to assess their own codebases for dependencies on languages like COBOL or Fortran. As the workforce of legacy language experts shrinks, institutions may need to integrate AI tools into their modernization strategies while maintaining rigorous testing protocols.
What to do and how to stay safe: OpenAI
- Audit your organization's software inventory to identify critical systems still running on legacy languages like COBOL, Fortran, or Visual Basic.
- Evaluate existing migration plans and consider how AI-assisted tools might reduce the cost and time required for transitioning to modern languages.
- Ensure that any AI-generated code undergoes strict expert review and testing against business rules before being deployed to production environments.
- Monitor industry benchmarks and vendor updates for improvements in AI models' ability to preserve specific program behaviors during translation.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Which OpenAI model performed best in migrating COBOL code to modern languages?
GPT-6.1 Sol achieved the highest accuracy in migration tests, reaching 97.9 percent with high reasoning effort when converting COBOL to languages like C++ or Java.
What percentage of global banks still use COBOL systems?
According to OpenAI, approximately 40 percent of the world's banks still rely on COBOL systems for their core operations.
How does OpenAI measure the effectiveness of its models in legacy languages?
The company uses an evaluation suite called MultiLangBench, which tests models on tasks such as creating SQLite-compatible database engines in specific programming languages.
