We are the Frontier
World

OpenAI Released GPT-6 Astra with Computer-Use Skills

The new model automates legacy banking interfaces, signaling a shift toward AI-driven software navigation.

Updated on Sept. 22, 2026 in Artificial Intelligence

Bold flat-color editorial illustration showing a brass gear beside a geometric prism, representing legacy software integration with modern AI.
OpenAI has released the GPT-6 Astra model, which utilizes computer-use capabilities to automate interactions within legacy banking systems and software. AI Illustration. Upload story photo >

Live Poll

Do you trust artificial intelligence to manage sensitive systems without human supervision?

OpenAI has released the GPT-6 Astra model, which features computer-use capabilities that allow it to read screens and execute tasks within programs. The model is currently available via the OpenAI API.

Why it matters

Financial institutions are increasingly testing these agents to interact with legacy banking software, potentially bypassing the need for costly and complex system replacements. This development offers a shortcut for banks currently reliant on mainframes and COBOL-based infrastructure.

On the OSWorld 2.0 benchmark, the GPT-6 Astra model achieved a 72.6% success rate, an improvement over the 65.7% score of the GPT-5.6 Sol model. The new model further reduced task completion time to 40 minutes per task, compared to the 75 minutes required by its predecessor.

The players

OpenAI

An artificial intelligence research organization and developer of large language models and autonomous agent technology.

Goldman Sachs

A global investment banking firm currently testing AI agents to automate trade accounting and client onboarding processes.

TSB Bank

A United Kingdom financial institution that provides a case study for the risks of legacy platform migrations.

The details

The GPT-6 Astra model functions as a computer-use agent that parses visual screen data to replicate human interaction with legacy software. Rather than relying on rigid, pre-programmed scripts, the model analyzes the graphical user interface to fill out forms, type text, or click buttons. This allows the AI to navigate older banking platforms that were never designed for modern API-based integration.

Timeline

  1. September 2026: OpenAI released the GPT-6 Astra model.

  2. June 2026: Benchmarking of AI agent performance was conducted.

  3. February 2026: The Goldman Sachs CIO discussed the firm's testing of AI agents.

  4. April 2018: TSB Bank migrated its platform.

  5. 1959: The COBOL programming language was introduced.

The Tech Race

Financial institutions are utilizing AI agents to modernize infrastructure, aiming to avoid the operational risks demonstrated by the 2018 TSB Bank platform migration. This effort reflects a broader industry movement to preserve COBOL-based mainframes that still handle 95% of U.S. ATM transactions.

The GPT-6 Astra model is available now for developers and enterprise users via the OpenAI API. This technology primarily targets banking and financial workflows, where it could eventually automate repetitive tasks currently handled by human operators on legacy systems.

The takeaway

Financial firms are betting that AI agents can bridge the gap between decades-old banking backends and modern service expectations. Watch for updates from major banks on the integration of these agents into core trade accounting and onboarding workflows in late 2026 and 2027.

Further reading

For broader trends in enterprise deployment, visit the Artificial Intelligence section.

Source note: This article includes information reported by PYMNTS.

Live Poll

Do you trust artificial intelligence to manage sensitive systems without human supervision?