OpenAI Introduces Major AI Breakthroughs to Modernize COBOL and Low-Resource Programming Languages

OpenAI has unveiled major advances in its artificial intelligence model capabilities for low-resource programming languages such as COBOL, Fortran, and Visual Basic. The new GPT-6 model series achieves up to 87.4% accuracy on multi-language software creation and 97.9% accuracy on legacy COBOL migration while drastically reducing operational API costs. These developments enable enterprise software teams to maintain and modernize mission-critical infrastructure that powers global banking, financial transactions, and public safety systems.

RMN Digital AI Desk
New Delhi | October 20, 2026

OpenAI Breakthroughs Target the Legacy Code Infrastructure Powering Global Enterprise Systems

Artificial intelligence development has historically favored widely spoken digital languages such as Python and JavaScript due to the massive volume of public code and online documentation available for training data. However, the world’s most critical infrastructure often runs on older, low-resource programming languages that receive far less representation in modern training datasets.

To bridge this gap, OpenAI has announced significant performance gains across its latest model family, demonstrating dramatic improvements in generating, debugging, and migrating code written in languages including COBOL, Fortran, Visual Basic, and OCaml.

The Strategic Importance of Low-Resource Languages

Low-resource programming languages underpin essential services worldwide. COBOL, a language first developed in 1959, still supports more than 40% of online banking systems, 80% of in-person credit card transactions, and 95% of ATM transactions globally. Similarly, Fortran remains a cornerstone of scientific computation, powering weather forecasting models and nuclear reactor safety analysis.

Also Read and Subscribe to:

[ Technophile | Tech Insights LinkedIn Newsletter ]

[ Movie Mate | Global Entertainment Intelligence LinkedIn Newsletter ]

Despite their vital role, maintaining and modernizing legacy codebases poses severe operational challenges for enterprises:

• Documentation Gaps: Original system documentation is frequently incomplete or outdated.
• Loss of Expertise: Original system developers have often retired or left organizations, leaving institutional knowledge unrecorded.
• Complex Business Logic: Critical business logic and specific operational rules exist solely within legacy source code.
• Strict Numeric Rules: COBOL relies on fixed-point arithmetic rules that discard excess decimal digits unless explicit rounding is requested. Small translation errors during code migration can alter financial outcome calculations without triggering software execution errors.

MultiLangBench: Building Databases from Scratch Across 20 Languages

To rigorously evaluate how well artificial intelligence models navigate low-resource languages, OpenAI established two complementary benchmarks. The first, titled MultiLangBench, tasks models with building a fully functional SQLite-compatible database engine from scratch across 20 distinct programming languages.

The evaluation tests an AI model’s capacity to manage software state, structure repositories, handle runtime errors, and add new features during extended coding sessions.

In benchmark tests:

• GPT-6 Astra achieved an average test pass rate of 87.4% across all 20 tested languages at maximum reasoning effort, representing a substantial leap from the 26.8% pass rate recorded by GPT-5.6 Sol.
• Language-Specific Gains: Astra achieved scores above 83% in 19 of the 20 languages tested, including 72.1% in COBOL (up from 10.9%), 85.8% in OCaml (up from 25.6%), 83.5% in Fortran (up from 21.4%), and 89.4% in Visual Basic (up from 27.5%).
• GPT-6 Luna demonstrated a budget-conscious alternative, reaching a 40.3% average test pass rate across all languages at an estimated average API cost of $0.77 per sample—approximately 78% lower than the $3.48 cost per sample associated with GPT-5.6 Sol.

COBOL Migration: Preserving Byte-for-Byte Output Parity

OpenAI’s second evaluation specifically tests code migration by requiring models to convert legacy COBOL programs into modern languages—including C++, Java, Kotlin, Python, and Rust—while preserving exact behavioral output. Success is evaluated against hidden validation suites that demand strict byte-for-byte parity on observable outputs.

Results from the COBOL migration evaluation highlight key technical strides:

• GPT-6.1 Sol achieved 86.6% migration accuracy at low reasoning effort, compared to a 5.4% accuracy rate for GPT-5.6 Sol under identical settings, while reducing overall API costs by 39%.
• High-Reasoning Performance: When evaluated at higher reasoning effort settings, GPT-6.1 Sol achieved an accuracy rate of 97.9%, bringing the COBOL migration benchmark close to full evaluation saturation.

Implications for Enterprise Modernization and Software Engineering

By combining higher task accuracy with significant reductions in computational cost, these advancements address long-standing hurdles in legacy system maintenance. Rather than replacing human oversight, these tools serve as force multipliers for enterprise engineering teams working to inspect legacy repositories, execute bug fixes, and migrate software safely.

OpenAI has open-sourced the MultiLangBench framework and test suite on GitHub, allowing independent developers and enterprise organizations to inspect benchmark parameters and evaluate model performance against their specific tech stacks.