OpenAI's AI Models Score 97.9% on COBOL Migration as Banks Face Legacy Code Crisis

OpenAI Targets Banking's Legacy Code Problem

OpenAI announced on October 19, 2026, that its AI models have achieved significant improvements in working with legacy programming languages including COBOL, Fortran, and Visual Basic1

1

2

2

. The development addresses a critical challenge facing financial institutions: over 40% of the world's banks still rely on COBOL systems, which also support 80% of in-person credit card transactions and 95% of ATM transactions3

3

. As COBOL programmers age out of the workforce, banks face mounting pressure to modernise legacy software or risk operational disruptions within the next few years.

GPT-6.1 Sol Achieves 97.9% Accuracy on COBOL Migration

OpenAI's GPT-6.1 Sol demonstrated breakthrough performance in COBOL migration tests, achieving 86.6% accuracy at low reasoning effort compared to just 5.4% for its predecessor GPT-5.6 Sol2

2

. At high reasoning effort, the model reached 97.9% accuracy, bringing the test close to saturation. The lower reasoning setting also delivered cost savings, operating at 61% of GPT-5.6 Sol's API cost. These AI models can now translate COBOL programs into C++, Java, Kotlin, Python, or Rust while preserving critical behaviors1

1

. The migration process requires byte-for-byte parity on observable outputs, and OpenAI considers a migration successful only if all tests pass3

3

.

MultiLangBench Reveals Dramatic Improvements Across Languages

To evaluate its models' capabilities, OpenAI created MultiLangBench, an AI evaluation suite that challenges models to build SQLite-compatible database engines from scratch in specified programming languages1

1

. GPT-6 Astra achieved an average test pass rate of 87.4% across 20 languages at maximum reasoning effort, a dramatic improvement from GPT-5.6 Sol's 26.8% pass rate2

2

. Astra scored above 83% in 19 languages, though COBOL remained the hardest to crack, with its score rising from 10.9% to 72.1%. The more affordable GPT-6 Luna averaged 40.3% with an estimated API cost of $0.77 per sample, approximately 78% less than GPT-5.6 Sol's $3.482

2

. OpenAI is publishing the MultiLangBench setup, test suite, and results for each language on GitHub.

Preserving Critical Banking Behaviors During Code Translation

The complexity of COBOL migration extends far beyond simple code translation. OpenAI's models must preserve specific behaviors such as how COBOL rounds numbers with excess decimal digits and handles missing records1

1

. COBOL's fixed-point arithmetic drops extra decimal digits unless rounding is explicitly requested, and a rewritten program that rounds differently can run without errors while producing incorrect financial results2

2

. Hidden tests require outputs to match originals byte for byte, ensuring financial calculations remain accurate. OpenAI acknowledges that maintaining and modernising legacy software is difficult because documentation may be incomplete, original developers may have left the company, and important business rules may only exist in code3

3

.

Expert Review Still Required for Production Systems

Despite the impressive accuracy rates, OpenAI emphasizes that applying these capabilities to production software still requires expert review and testing against the system's business rules and dependencies1

1

3

3

. Programs handling banking tasks demand the highest accuracy standards. The company's work on improving COBOL capabilities was guided by feedback from developers and enterprise customers who reported that coding models struggle with low-resource languages—those with less available code and documentation for model training compared to widely used languages like Python and JavaScript2

2

. COBOL, created in 1959, has climbed from 23rd place in October 2025 to 16th in the TIOBE popularity index, reflecting continued enterprise dependence even as the pool of engineers able to maintain these systems shrinks. Watch for banking institutions to begin piloting these AI-assisted migrations as they race to modernise before the COBOL programmer shortage becomes critical.