Editorial illustration for OpenAI says GPT-6 improves legacy programming for COBOL and Fortran
GPT-6 Masters COBOL and Fortran Legacy Code
COBOL turns 65 next year, and OpenAI wants GPT-6 to speak it fluently. The company says its newest models show sharp gains on legacy programming languages, the kind of code most engineers never learn but that still runs the machinery of daily life. IBM puts a number on that reliance: COBOL handles more than 40 percent of online banking systems and 95 percent of ATM transactions worldwide. Fortran and Visual Basic sit in the same bind, widely deployed but starved of the training data that modern languages like Python enjoy in abundance.
To measure whether GPT-6 actually closes that gap, OpenAI built two tests. One, called MultiLangBench, asks models to construct a SQLite-compatible database engine from scratch across 20 languages. The other focuses narrowly on COBOL migration, checking how well a model can translate decades-old code into something like Rust or Kotlin without breaking it.
Banks and government agencies have spent years quietly trying to solve this translation problem, often by paying for scarce human expertise. OpenAI's numbers, laid out below, suggest its models are starting to close that gap on their own.
OpenAI improved its latest AI models for "low-resource" languages like COBOL, Fortran, and Visual Basic. These languages have far less available training code than something like Python, but according to OpenAI, they prop up critical infrastructure worldwide. IBM estimates that COBOL powers more than 40 percent of online banking systems and 95 percent of all ATM transactions.
Why this matters
For banks still running COBOL mainframes, an 87.4 percent pass rate on a synthetic database-building task is interesting but not a green light to start handing production code to GPT-6. These systems move money for hundreds of millions of people, and IBM's own numbers on ATM transactions make clear how little room there is for error. What's notable here is OpenAI building a dedicated benchmark for "low-resource" languages at all, a tacit admission that model quality has been uneven outside Python and JavaScript for years, and that enterprises with Fortran and Visual Basic codebases have been getting a worse product without much acknowledgment.
If GPT-6 actually closes that gap, it opens real work for teams maintaining legacy infrastructure that most new engineers have no training in and no interest in learning. We'd want to see independent testing against real production code, not just a purpose-built eval, before treating this as settled. The more immediate question for founders building dev tools: does this make COBOL migration or maintenance products viable businesses, or does OpenAI's own benchmark data mean that market shrinks before it ever grows.
Related Reading
Browse all OpenAI news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source
Related Topic OpenAI Developer: AI Agent "Swarms" Waste Tokens, Create "Coordination Tax
AI Tools & Apps
From Archives Purdue, MIT researchers develop generative AI robotic system for stroke therapy
From Archives Astrolab's FLIP rover will run the first custom-built LLMs on the moon.
From Archives OpenAI Acquires Orchestra Bio to Expand AI Agents for Biotech Drug Discovery
Business & Startups
From Archives Mathematicians Criticize OpenAI's Navier-Stokes Proof Release
Open Source
Common Questions Answered
Why does OpenAI focus on improving GPT-6's performance with legacy programming languages like COBOL and Fortran?
Legacy languages like COBOL and Fortran have far less available training data compared to modern languages like Python, making them low-resource languages that AI models struggle with. Despite their scarcity in training data, these languages power critical global infrastructure, with COBOL handling over 40 percent of online banking systems and 95 percent of ATM transactions worldwide. OpenAI's focus on improving GPT-6 for these languages addresses a significant gap in AI capabilities for maintaining and updating mission-critical systems.
What percentage of global ATM transactions does COBOL currently handle according to IBM?
According to IBM's estimates cited in the article, COBOL handles 95 percent of all ATM transactions worldwide. This statistic underscores the enormous reliance on legacy COBOL systems for daily financial operations globally.
What does OpenAI's 87.4 percent pass rate on synthetic database-building tasks mean for production banking systems?
While the 87.4 percent pass rate demonstrates significant improvement in GPT-6's ability to handle COBOL tasks, it is not considered a green light for deploying the model to generate production code for banking systems. Given that these systems handle money for hundreds of millions of people and the critical nature of ATM infrastructure, there is minimal room for error, meaning additional validation and human oversight would be necessary before relying on AI-generated COBOL code in production environments.
What is significant about OpenAI building a dedicated benchmark for low-resource programming languages?
Creating a dedicated benchmark for low-resource languages like COBOL, Fortran, and Visual Basic represents a tacit admission that model quality has been uneven across different programming languages. This focused approach acknowledges that legacy languages require specialized attention and measurement to ensure AI models can effectively support the infrastructure that still depends on them.
Further Reading
- Advancing coding in low-resource programming languages - OpenAI
- OpenAI says GPT-6.1 Sol scores 86.6% on its COBOL migration test, up from 5.4% - ThinkFacility
- OpenAI looks to help banks modernise legacy software their operations still depend on - The Economic Times
- OpenAI says GPT-6.1 Sol is dramatically better at migrating legacy code - BleepingComputer
- OpenAI Launches a Coding Model for Older Programming Languages - Thurrott
