1. What did OpenAI announce about COBOL and other legacy languages?
In an OpenAI COBOL update posted on X on October 19, 2026, the company said that the latest improvements to its model make it better at writing and migrating code in COBOL, Fortran and Visual Basic, languages the company groups under the label "low-resource languages." The company framed the work around two aims: helping teams understand business logic written decades ago, and preserving the way existing systems behave while they are modernized.
OpenAI also asked developers and enterprise customers which other languages or workflows should receive similar attention, an indication that the effort is meant to widen. The announcement was brief, and the company's claims about performance come from its own testing rather than independent evaluation.
The update targets a part of the software world that rarely makes headlines but underpins much of the financial system. Banks, insurers, government agencies and large corporations still run critical processes on code that predates the modern web, and the legacy code modernization market has become a competitive target for AI developers.
2. What does "low-resource" mean for a programming language?
In machine learning, a low-resource language is one with relatively little publicly available training material. Python and JavaScript are the opposite: millions of open repositories, tutorials and forum answers give models abundant examples to learn from. COBOL, Fortran and Visual Basic have far less open code, because much of what exists lives inside private corporate systems that were never published.
Research on code-generation models has consistently found that they perform worse on such languages than on mainstream ones. Academic work has attributed COBOL's difficulty to the shortage of open material, and has noted that Fortran faces a similar problem. Benchmarks published by researchers have reported pass rates and compilation rates ranging from single digits to roughly 40 percent for COBOL tasks on general-purpose models.
That history explains why a vendor claiming progress here draws attention. Closing even part of the gap would make AI assistance usable in environments where it has so far been unreliable.
3. Why does COBOL still matter to banks and governments?
COBOL, created in 1959, was designed for business data processing, and it excels at handling large volumes of financial records. Estimates of how much of it remains in production run to hundreds of billions of lines, powering functions such as account processing, payments, benefits administration and insurance claims.
Replacing these systems is difficult because they work. They have been refined over decades, and they encode business rules that may not be written down anywhere else. A rule about how interest is rounded or how a particular account type is treated may exist only in the code, and the people who originally wrote it may have retired.
This is why OpenAI emphasized understanding decades-old logic as a goal. Before code can be safely translated, someone or something must work out what it actually does, including its edge cases.
4. How does the modernization challenge work in practice?
Migrating legacy code usually involves several stages: reading and documenting the existing programs, mapping their dependencies, translating them into a newer language or architecture, and testing that the new version produces identical results. Each stage consumes engineering time, and the last is often the most demanding because it requires a complete set of test cases.
Organizations face rising maintenance costs and a shrinking pool of engineers who know these languages. As experienced COBOL and Fortran specialists retire, the knowledge needed to maintain critical systems becomes scarcer and more expensive, which pushes companies toward either costly rewrites or continued dependence on a dwindling workforce.
AI tools promise to speed the early stages in particular, by explaining unfamiliar code and drafting translations that humans review. The promise of lower modernization costs is the commercial reason the major AI developers are competing for this work.
5. Who else is competing in legacy code?
OpenAI is not alone. Anthropic has announced its own COBOL-related updates earlier in 2026, and IBM, which has long sold mainframe systems, offers watsonx Code Assistant for translating COBOL into Java. Cloud providers and consulting firms also market migration services built around AI.
The competitive picture matters for customers because it affects pricing and the choice of tools. A bank evaluating options may consider a model vendor, a mainframe specialist or an integrator, each with different strengths: model vendors offer general capability, while established mainframe companies bring knowledge of the environments in which the code runs.
For Fortran, which is widely used in scientific and engineering computing, the use case differs. There, the goal is often to translate and optimize numerical code so it runs efficiently on modern hardware, rather than to preserve business rules.
6. Why is human oversight still essential?
Even a much-improved model can make mistakes that are hard to see. In regulated industries, a migration that changes behavior, even slightly, can cause compliance problems. A rounding difference in a payment calculation, repeated across millions of transactions, can produce material errors and attract regulators' attention.
For that reason, specialists stress verification: running old and new systems side by side, comparing outputs on large test sets, and keeping engineers responsible for sign-off. AI output is best treated as a draft that speeds up skilled people rather than as a replacement for review.
OpenAI's own wording points in that direction. By highlighting the preservation of existing system behavior as a goal, the company acknowledges that equivalence, and not just working code, is the standard that enterprises will apply.
7. What makes legacy code so costly to maintain?
Cost accumulates in several ways. Older systems often depend on specialized hardware and licensing, documentation tends to be incomplete, and every change carries the risk of unintended side effects in programs that interact in ways nobody fully mapped. Companies frequently pay a premium for the few engineers able to work on them.
There is also an opportunity cost. Time spent keeping old systems running is time not spent building new products, and rigid back-end systems can slow a firm's ability to launch services such as instant payments or new account types. This is why modernization appears on the agenda of many chief technology officers even when the existing systems are technically reliable.
The risk of doing nothing grows with each retirement of a veteran engineer, while the risk of acting too quickly is that a flawed migration disrupts services customers depend on. Tools that reduce the effort of understanding old code address the first risk without necessarily raising the second.
8. What would success look like for customers?
For an enterprise, a useful model would explain unfamiliar programs accurately, produce translations that pass existing tests without modification, and flag the places where it is uncertain instead of guessing. Measured improvement in these areas would reduce the hours engineers spend on routine reading and conversion.
Cost is the second yardstick. If agent-based workflows can translate and optimize code at a modest price per unit of work, projects that were previously uneconomic, such as converting a large library of scientific routines, could become feasible. Customers will compare that expense with the cost of specialist consultants and with the review effort still required.
Third is trust. Firms in finance and government will want audit trails showing how a translation was produced and tested. A model that cannot document its work may still be useful for drafting, but it will struggle to win approval for production changes.
9. Why does this announcement matter beyond software teams?
Many of the systems in question sit at the center of financial infrastructure, so the pace at which they can be modernized affects how quickly banks can adopt new payment rails, real-time settlement and other services. Faster migration could, in principle, narrow the gap between legacy institutions and newer, cloud-native competitors.
At the same time, the announcement illustrates how AI developers are moving from general chat tools toward specific, high-value enterprise problems. Legacy modernization is attractive because the work is repetitive, the buyers have budgets, and the pain of the status quo is well understood.
The same logic applies to the customers' own governance. A bank adopting AI-assisted migration will need clear policies on who reviews generated code, how test coverage is measured and when a project is allowed to move from a pilot to production, since those procedures determine whether faster tools translate into safer outcomes.
Observers will therefore treat results from this segment as an early indicator of how effectively AI models can handle complex, high-stakes work, where an error is not merely inconvenient but potentially expensive.
10. What happens next?
Because the performance claims are the company's own, the next step is independent testing. Researchers and customers will want to compare the model on realistic legacy programs, with their idiosyncratic dialects and undocumented business rules, rather than on tidy benchmark tasks.
OpenAI's request for feedback suggests that additional languages and workflows may follow, though it has not named any or given a timetable. Enterprises considering the technology will likely begin with pilots on non-critical systems, measuring accuracy and review effort before committing larger migrations.
Readers should also keep the limits of the announcement in mind. OpenAI described its goals and invited feedback, but a statement on social media is not a technical report, and the details of how the model was evaluated, on which programs and against which baselines, will determine how much weight the claims deserve. Until independent results appear, cautious optimism is the reasonable stance for technology leaders.
The announcement is best read as one more move in a crowded field. Whether it reduces modernization costs in practice will depend on results in real organizations over the coming months.