OpenAI just released GPT-5.4, which adds native computer control to their most capable reasoning model. The system can operate desktops through screenshots and mouse commands, achieving 75% accuracy on OSWorld desktop tasks compared to 72% for humans. This is OpenAI’s first general-purpose model with built-in computer use capabilities.
The professional work improvements are substantial. GPT-5.4 matches or beats industry professionals in 83% of knowledge work comparisons across 44 occupations, which is up from 71% for GPT-5.2. On spreadsheet modeling tasks that junior investment bankers perform, the model scores 87% versus 68% for the previous version. Human evaluators preferred GPT-5.4’s presentations 68% of the time due to better aesthetics and visual variety.
GPT-5.4 also introduces tool search, which dramatically reduces token costs for agent workflows. Instead of loading all tool definitions upfront, the model looks up specific tools when needed. This approach cut token usage by 47% on Scale’s MCP Atlas benchmark while maintaining the same accuracy. For systems with thousands of available tools, the efficiency gains are material.
The computer control works through Playwright code generation and direct screenshot interaction. GPT-5.4 can navigate websites, interact with desktop applications, and complete multi-step workflows across different software environments. The model supports up to 1 million tokens of context, enabling longer planning horizons for complex tasks.
I’m interested in the enterprise adoption timeline here. Computer control creates genuine productivity value, but it also introduces new security and compliance considerations. We will need to evaluate which workflows justify giving AI systems direct access to their software environments. The 95% success rate on property tax portals that Mainstay reported suggests the reliability is approaching production readiness for specific use cases.
OpenAI priced GPT-5.4 at $2.50 per million input tokens, up from $1.75 for GPT-5.2. The Pro version costs $30 per million input tokens. Despite higher per-token pricing, the improved efficiency means lower total costs for many workflows. The model uses significantly fewer tokens to solve the same problems compared to earlier versions.
The release consolidates OpenAI’s reasoning and coding capabilities into a single model that can operate computers independently. Professional services firms testing GPT-5.4 report it excels at creating slide decks, financial models, and legal analysis while running faster than competitive models. This positions OpenAI directly against specialized business AI tools that focus on specific professional workflows.
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.