OpenAI says its newest AI models can translate decades-old bank code into modern languages, which could speed up upgrades behind ATMs, though the proof is its own tests.
OpenAI says its latest AI models have gotten much better at rewriting very old software into newer programming languages. That matters because much of the code running banks, government agencies and other critical services is decades old. The company announced the results on Monday. They come from tests OpenAI built itself.
COBOL was invented in 1959 and still handles about 95% of all ATM transactions. Fortran, from 1957, is still widely used to simulate how nuclear reactors work. Many engineers who know these languages are retiring and taking their skills with them.
Meanwhile, younger programmers lean on AI assistants like Codex and Claude Code. The data behind those tools is not public, but they were very likely trained mostly on newer languages. That leaves a widening gap between how new code gets written and the software that runs much of finance, IT services and government.
OpenAI created two new tests, called benchmarks, to see whether its models could help close that gap. It says its work was guided by feedback from developers and enterprise customers.
The second test is tricky. Financial systems follow different rules, such as how and when decimals get rounded. A model has to catch that context, not just swap words, much like translating a novel while keeping its meaning and tone.
OpenAI is not alone. Earlier this year, Anthropic wrote that its Claude Code could help businesses update COBOL systems, a process that has been expensive and slow. It said teams could finish in quarters instead of years. Bloomberg reported that IBM shares fell 13% after that post, the biggest one-day drop since 2000. IBM mainframes are widely used to run COBOL.
The scores come from OpenAI's own benchmarks, and the company is grading its own work. The material does not say how the models perform on real bank systems, which institutions plan to use them, or how much the work would cost.
A near-perfect score on a test is not the same as a safe switch for a system that moves people's money. Rounding errors, for example, could matter a great deal at scale. For now, this is a promising claim, not a proven fix.
The code behind your bank is old and AI may speed up upgrades, but the evidence so far comes from OpenAI's own tests.
Google's Gemini chatbot can now open and edit some files in the chat, PDFs included. A reviewer found it handy for quick fixes but not a replacement for Docs or Acrobat.
Read more: Gemini can now edit your PDFs and Google files inside the chatA tech writer fed his Windows error logs to ChatGPT and learned his encrypted drive had no recovery key, a problem that could have locked him out for good.
Read more: ChatGPT read his laptop's error logs and found three hidden problemsA TikTok Shop affiliate says a strong first line, an honest demo and steady posting earned him six figures last year. Here is how it works and what it costs to try.
Read more: How a TikTok creator earns six figures with a 3-second opening line
