OpenAI GPT-6 Achieves 97.9% Accuracy on COBOL Migration Benchmark

OpenAI announced significant performance improvements in its latest artificial intelligence models for handling legacy programming languages. According to reporting by The Decoder, the company’s newest models demonstrate substantially higher accuracy in translating and building software in older languages such as COBOL, Fortran, and Visual Basic. This development addresses challenges in maintaining critical infrastructure that relies on these outdated but pervasive coding standards.

The performance gains are quantified through two specific evaluations conducted by OpenAI. On the MultiLangBench, which requires models to construct a SQLite-compatible database engine from scratch, GPT-6 Astra achieved an average test pass rate of 87.4 percent across twenty languages. This represents a major increase from the approximately 38 percent pass rate recorded by the previous GPT-5.6 Terra model. The benchmark tests the ability to build complex software structures in languages with limited available training data.

Context and Background

OpenAI categorized COBOL, Fortran, and Visual Basic as "low-resource" languages because they have far less publicly available code for training compared to modern languages like Python. Despite their age, these languages support essential global infrastructure. IBM estimates that COBOL powers more than 40 percent of online banking systems and handles 95 percent of all ATM transactions worldwide. This widespread reliance makes accurate modernization tools critical for financial institutions seeking to update their technology stacks without disrupting services.

The second evaluation focused specifically on COBOL migration tasks. This benchmark measures the accuracy of translating COBOL code into modern languages including C++, Java, Kotlin, Python, or Rust. Under low reasoning effort settings, GPT-6.1 Sol achieved 86.6 percent accuracy, a dramatic improvement over the 5.4 percent accuracy of GPT-5.6 Sol. When the model applied maximum reasoning effort, accuracy climbed to 97.9 percent. These results suggest a potential pathway for automating the tedious process of modernizing legacy financial software.

Who Is Affected

Financial institutions and government agencies that maintain legacy systems are the primary stakeholders in these developments. As Threat Times reported on October 20, 2026, OpenAI models showed improved capabilities for migrating legacy COBOL banking systems, addressing critical workforce gaps. The current data reinforces those earlier findings with specific metrics on model performance. Organizations responsible for maintaining decades-old codebases may find these tools useful for accelerating modernization efforts, though the high accuracy rates depend on specific model configurations and reasoning settings.

What should you do?

Security and technology professionals should monitor vendor announcements regarding AI-assisted code migration tools. Teams managing legacy infrastructure should evaluate whether automated translation tools can reduce manual effort in modernization projects. It is important to verify that translated code meets security standards and functions correctly in production environments. Organizations should also assess the cost implications of using high-reasoning models versus lower-cost alternatives for different migration tasks.

What to do and how to stay safe: OpenAI

• Review current legacy codebases to identify critical systems that rely on COBOL or Fortran.
• Evaluate the accuracy requirements for your specific migration projects against the reported benchmarks.
• Test automated translation outputs in isolated environments before deploying to production systems.
• Monitor for updates on model versions that offer better cost-to-performance ratios for large-scale tasks.

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

How does GPT-6 compare to previous models on legacy language tasks?

GPT-6 Astra achieved an 87.4 percent pass rate on MultiLangBench, up from about 38 percent for GPT-5.6 Terra, while GPT-6.1 Sol reached 97.9 percent accuracy on COBOL migration.

Which programming languages showed the most improvement?

OpenAI highlighted improvements in low-resource languages such as COBOL, Fortran, and Visual Basic, which have less available training code than modern languages.

What is the scope of COBOL usage in financial systems?

IBM estimates that COBOL powers more than 40 percent of online banking systems and 95 percent of all ATM transactions globally.