Replacing legacy banking systems often takes five years and costs hundreds of millions of dollars. Many financial institutions still rely on COBOL-based cores, which handle approximately 95% of U.S. ATM transactions. According to DXC data, over 40% of banks continue to use these systems, and 45 of the 50 largest global banks depend on mainframes for critical operations.
OpenAI introduced the GPT-6 Astra model, which can independently operate computer interfaces by reading screens, filling out forms, and navigating between programs. In internal tests using an offline version of OSWorld 2.0, Astra achieved a score of 72.6% with an average task time of 40 minutes. Its predecessor, GPT-5.6 Sol, scored 65.7% and required 75 minutes per task.
Goldman Sachs has been testing AI agents for trade accounting, transaction processing, and client onboarding. Chief Information Officer Marco Argenti noted that the bank spent six months collaborating with Anthropic engineers to develop these digital tools. The bank is exploring these agents to handle complex, labor-intensive workflows and potentially reduce reliance on external vendors.
While AI agents show promise, current technology faces limitations. On the full OSWorld 2.0 benchmark, which consists of 108 tasks, the top-performing agent successfully completed approximately 31% of jobs from start to finish. Researchers highlight that long, realistic tasks present safety risks, and developing fully reliable computer-use agents remains an open research challenge.
