株式会社オブライト
AI2026-03-19

GPT-5.4 Computer Use API — Enterprise Automation Goes Mainstream

Released on March 5, 2026, GPT-5.4 Computer Use API brings native screen control to general-purpose AI models. With 75% on OSWorld-V benchmark surpassing human performance at 72.4%, it's accelerating practical automation for legacy systems and RPA replacement.


Why GPT-5.4 Computer Use API Matters

On March 5, 2026, OpenAI launched GPT-5.4 Computer Use API, the first general-purpose AI model with native screen control capabilities. While AI automation has traditionally relied on API integrations or script generation, GPT-5.4 enables direct browser control via Playwright, along with mouse and keyboard automation. This breakthrough allows AI agents to interact with legacy systems and GUI applications just as humans do, opening up automation opportunities previously limited to inflexible RPA tools. For businesses dealing with non-API systems or complex screen workflows, GPT-5.4 offers unprecedented flexibility and accuracy.

Surpassing Human Performance at 75% on OSWorld-V

GPT-5.4 Computer Use API achieved a 75% score on the OSWorld-V benchmark, exceeding human performance at 72.4%. This benchmark evaluates multi-step task execution in real operating system environments, including browser operations, file management, and cross-application workflows. Notably, GPT-5.4 reduced individual assertion errors by 33% and overall errors by 18% compared to GPT-5.2. These improvements in reasoning accuracy enable reliable automation even for complex business flows involving conditional branching and exception handling. The model's robustness makes it suitable for production deployment in enterprise environments.

1M Token Context and Reasoning Level Optimization

The API version of GPT-5.4 supports a 1 million token context window, allowing batch processing of large documents and log files. Additionally, it offers five reasoning levels (none/low/medium/high/xhigh) to optimize cost and accuracy based on task complexity. For example, simple data entry tasks can use 'low' reasoning, while contract analysis or decision-making tasks benefit from 'high' or 'xhigh' settings. Pricing is set at $2.50 per 1M input tokens and $15 per 1M output tokens (Pro version: $30/$180), competitive with traditional RPA licensing costs while offering superior capabilities.

Tool Search: 47% Token Reduction

GPT-5.4 incorporates Tool Search functionality, which achieved a 47% token reduction across 250 MCP Atlas tasks. Traditionally, AI agents had to load the full list of available tools and APIs into context for every request. Tool Search dynamically retrieves only the necessary tools, preventing context bloat even when hundreds of APIs are available in complex enterprise environments. This feature significantly improves response speed and cost efficiency, making it essential for large-scale production deployments where multiple systems and services need to be orchestrated.

Playwright Integration for Browser Automation

GPT-5.4 integrates seamlessly with the Playwright library, enabling complete automation of browsers like Chrome and Edge. Use cases include web form data entry, multi-page workflow execution, scraping, and test automation. Unlike traditional RPA tools that break easily when UI layouts change, GPT-5.4 combines visual recognition with natural language understanding to handle minor UI modifications gracefully. This robustness is particularly valuable for automating legacy internal systems and SaaS admin panels, drastically reducing implementation and maintenance costs.

RPA Replacement and Test Automation Applications

GPT-5.4 Computer Use API excels at unstructured tasks and workflows requiring judgment—areas where traditional RPA tools struggle. For instance, it can handle end-to-end processes like extracting invoice data, entering it into systems, and triggering approval workflows across multiple applications. In software UI test automation, QA engineers can describe test scenarios in natural language, and GPT-5.4 automatically generates and executes screen operations and assertions. This shifts QA focus from writing test code to test design. With superhuman accuracy and flexibility, the scope of practical automation is expanding rapidly.

Conclusion — Partnering for Production AI Automation

GPT-5.4 Computer Use API elevates AI automation from proof-of-concept to production-ready solutions. It unlocks automation for legacy systems, RPA replacement, and test automation—domains that previously required manual effort. Oblight Corporation offers business process design, system integration development, and deployment support leveraging GPT-5.4. If your organization seeks to implement AI automation while preserving existing systems, we invite you to consult with us. We optimize workflows and deliver measurable productivity gains tailored to your business needs.

Feel free to contact us

Contact Us