OpenAI GPT-5.4 officially released: AI evolves from 'conversational tool' to 'real computer operator' On March 5th, OpenAI officially launched GPT-5.4 (including Thinking/Pro version), which is their first universal flagship model with native support for computer use. 1. Core breakthrough: A qualitative change in computer control ability In the OSWorld Verified real desktop task test, the success rate of GPT-5.4 was * * 75% * *, exceeding the human average level (72.4%) for the first time, while the previous generation GPT-5.2 was only 47.3%. Actual effect: It can look at the screen, click the mouse, type the keyboard, and complete processes across applications like a real person. Uncle Lan (@ LufzzLiz) personally tested: Let it set a schedule on Mac, write a diary with Obsidian, and "get it done in minutes" without any additional skills throughout the process. 2. Comprehensive upgrade of professional work ability -Knowledge work benchmark GDPval AA: tied or won with 44 professional professions by * * 83% * * (previous generation 70.9%) -Internal evaluation of investment banking modeling: Jumping from 68.4% to * * 87.3%** -Generate presentation: 68% probability of manual blind testing better than previous generation -Programming: Inheriting GPT-5.3-Cdex capabilities, SWE Bench Pro scores 57.7% with lower latency 3. Practical revolutionary level optimization -Tool Search on-demand calling tool: When there are multiple tools, token consumption is reduced by * * 47% * *, and both cost and latency are reduced -Real time display of thinking process (Thinking mode): Users can intervene and adjust direction midway -Visual and document parsing greatly enhanced, supporting high-resolution 'original' image details -Overall more token efficient, faster and lower cost for the same task 4. Pricing and availability API input is $2.50/output is $15 (slightly increased from the previous generation), ChatGPT Plus/Team/Pro users have switched to GPT-5.4 Thinking by default, and the old version will be retired on June 5, 2026. The Pro version specializes in extremely complex tasks. Rational comparison with Claude Opus 4.6: GPT-5.4 leads in computer control, daily agent processes, and cost-effectiveness; Claude Opus 4.6 still has advantages in extreme coding (slightly higher in SWE Bench rankings) and deep multi-step reasoning. Conclusion: Dual holding is optimal - GPT-5.4 is the main focus, with Claude assisting in tackling hardcore code. GPT-5.4 is not just another chat model, but the first step towards truly turning AI into a 'digital employee'. It has turned 'just think about it' into 'just help you finish it', directly raising the efficiency ceiling for local agents, productivity tools, and Mac users again. Have you already mastered GPT-5.4? -I tried computer control and it was so enjoyable that it made me fly -Still watching Claude -Dual open mixed use -Pure cloud/local deployment party Forward to all friends interested in AI productivity - whoever uses it first will seize the efficiency dividend of 2026! GPT54 Computer Use AgenticiAI OpenAI Productivity Revolution MacStudio