OpenAI Hands Banks a Shortcut Around Billion-Dollar Core Rewrites

In a bank’s back office, an employee pulls up a customer’s account on a decades-old terminal, reads the balance and types it into a newer web form to finish the file. The two systems don’t talk to each other. The person in the chair is the connection.

That job exists because of how bank technology was built. The systems that hold balances, process loans and move payments are often 30 or 40 years old. They were written before software had standard ways to exchange data, so newer tools can’t plug into them directly. A bank that wants artificial intelligence (AI) in those workflows has had two choices. Replace the old system—which takes years—or build a custom connection for each one, which is slow and expensive.

A third option is emerging. OpenAI said that its new GPT-6 Astra model can operate a computer on its own. It reads the screen, fills out forms, updates records and moves between programs by clicking and typing, as an employee would. On an offline version of OSWorld 2.0, a test built to measure that skill, Astra scored 72.6% at about 40 minutes per task in OpenAI’s own testing. Its predecessor, GPT-5.6 Sol, scored 65.7% at about 75 minutes.

Old Cores Still Carry the Money

Systems written in COBOL, a programming language from 1959, still run an estimated 95% of U.S. ATM transactions, and hundreds of billions of lines of the code remain in production across finance, airlines and government, PYMNTS reported. Estimates from IT services company DXC, cited by Amazon Web Services, show that more than 40% of banks still run COBOL-based cores and 45 of the world’s 50 largest banks rely on mainframe applications for critical operations.

Replacing these systems is slow, costly and risky. Commonwealth Bank of Australia’s core replacement took five years and more than $1 billion Australian dollars, iTnews reported in 2012. When U.K.-based bank TSB moved to a new platform in April 2018, many of its 5.2 million customers lost access to banking services. The U.K.’s Financial Conduct Authority and Prudential Regulation Authority fined the bank 48.65 million pounds in 2022 and said it took until eight months for customers to regain normal access.

Banks have tried automating old screens before. Software robots, known as robotic process automation, follow fixed scripts. When a button moves or a screen changes, they stop working.

Agents Take the Employee’s Seat

Computer-use AI works differently. It looks at the screen, decides what to do and adjusts when something changes. OpenAI, Anthropic, Google and Amazon each now offer this kind of agent, No Jitter noted. For a bank, an agent could take the employee’s seat, pull the record from the terminal and enter it in the newer system without new plumbing.

Goldman Sachs is testing the idea on labor-intensive work. The bank spent six months with Anthropic engineers embedded in its teams to build agents for trade and transaction accounting and for client vetting and onboarding, Chief Information Officer Marco Argenti told CNBC in February. He called them “a digital co-worker” for jobs that are large, complex and process heavy. Argenti said Goldman was surprised at how well the model handled accounting and compliance work, and that the bank could drop some outside vendors as the technology matures.

Benchmarks Stop Short of the Back Office

A test score is not a working bank process. OSWorld 2.0 was designed to measure long jobs. Its 108 tasks take a skilled person a median of about 1.6 hours each. As of late June, the best agent had finished about 31% of those tasks from start to end, according to the leaderboard hosted by Snorkel AI. OpenAI’s 72.6% came on a smaller subset of the test that runs without internet access and awards partial credit, the company noted in its release.

The test’s authors found that long, realistic tasks expose safety risks that rarely appear in short ones. They called a trustworthy computer-use agent an open research problem.

GPT-6 Astra is already live in OpenAI’s application programming interface (API) with computer use among its supported tools, the company’s model page shows.