Desktop banking software
AI Models & Assistants / Knowledge library
computer-use-preview
Specialized model for computer use tool
Why learn computer-use-preview?
Offers the documented OpenAI model capabilities through hosted.
The problem it solves
Provides model inference for supported text / image inputs.
Where it came from
computer-use-preview: Specialized model for computer use tool Published aliases: computer-use-preview-2025-03-11. Input: text, image. Output: text. Checked 2026-10-13. Model capabilities are distinct from tools available in a particular agent.
Build your understanding
Work through these exercises in order. Try the task before opening a reference.
- Explain what computer-use-preview does.
- Identify a use case for computer-use-preview.
- Describe a limitation of computer-use-preview.
- Name a compatible tool for computer-use-preview.
- Compare computer-use-preview against an alternative.
Documentation and sources
- Primary source / model card (developers.openai.com)
Catalog research notes
- See the AI Workbench for source dates, pricing evidence and configuration details.
Connect the ideas
Built on
Real World Example
Automated Legacy Financial Reconciliation
Mid-sized accounting firms managing legacy desktop-based banking software that lacks modern APIs.
The business problem
Manual data entry from proprietary desktop banking interfaces into cloud-based ERP systems is error-prone and labor-intensive, creating significant operational bottlenecks during month-end closing.
How this topic solves it
The computer-use-preview model enables the automation of these legacy interfaces by interpreting visual screen data and executing mouse and keyboard inputs directly, effectively bridging the gap between non-API legacy software and modern data workflows.
- Configure the environment to host the legacy desktop application in a virtualized, headless display session.
- Initialize the computer-use-preview model within the Responses API to process visual snapshots of the banking interface.
- Define tool-use functions that map model-generated coordinates and keystrokes to specific UI elements like buttons or input fields.
- Implement a control loop that captures screen state, sends it to the model, and executes the returned tool calls to perform reconciliation.
- Validate the output by comparing the final screen state against expected ledger balances before committing data to the cloud ERP.
The flow shows how the model interacts with the desktop environment to perform tasks.
Visual snapshot of UI
Model processing visual state
Mouse/Keyboard actions
Final data destination
- Legacy App → renders UI → Screen Capture
- Screen Capture → sends visual input → computer-use-preview
- computer-use-preview → triggers action → Tool Execution
- Tool Execution → modifies state → Legacy App
- Legacy App → updates ledger → Cloud ERP
Expected business outcome
A reliable, automated pipeline that mimics human interaction with legacy software, reducing manual reconciliation time and minimizing transcription errors.
Limitations and checks
The model is constrained by a 1,024 max output token limit per response, which may truncate complex multi-step interaction sequences; developers should check for incomplete tool execution sequences by monitoring the finish_reason in the API response.