OpenAI has released GPT-5.4, a frontier model featuring autonomous computer-use capabilities and a 1-million token context window for professionals.
OpenAI has officially launched GPT-5.4, its most capable model designed for professional environments, now accessible via ChatGPT, the API, and Codex. Alongside the standard version, the company introduced GPT-5.4 Pro to cater to users requiring maximum computational power for highly complex tasks. The new iteration integrates the advanced coding foundation of GPT-5.3-Codex with enhanced reasoning and agentic workflows.
Native Computer Use and Advanced Efficiency
A defining characteristic of GPT-5.4 is its native computer-use capability. This allows the model to operate desktop environments autonomously by executing keyboard and mouse commands based on visual screen inputs. Supporting a massive 1-million token context window, the model secured a 75.0% success rate on the OSWorld-Verified benchmark. Furthermore, the introduction of a “Tool Search” function minimizes token consumption by dynamically locating and applying the correct software tools only when required, reducing operational costs and latency.
Industry Leaders Report Measurable Gains
Early adopters across various industries report significant performance improvements. Brendan Foody, CEO of Mercor, stated that the model excels in generating extensive deliverables like financial models while operating faster and more cost-effectively than competitors. In the legal sector, Niko Grupen, Head of Applied Research at Harvey, noted that the model achieved a 91% accuracy rate in complex contract analysis. Dod Fraser, CEO of Mainstay, confirmed a 95% first-attempt success rate across thousands of property tax portals, executing tasks three times faster.
Software development platforms are also leveraging the upgrade. Lee Robinson, VP of Developer Education at Cursor, mentioned that their engineers find the model highly assertive in handling ambiguous problems. Wade, CEO of Zapier, described it as the most persistent model available for executing multi-step workflows. OpenAI also confirmed that factual errors have been reduced by 33% compared to previous iterations.
