OpenAI launched GPT-5.2 on December 12th, positioning it as the most suitable AI model series for "professional knowledge work." The official statement indicates that GPT-5.2 can directly generate practical work results such as presentations, spreadsheets, code, and long document analysis, and supports cross-tool, multi-step task processing. It can also create front-end web games and greeting cards. According to feedback, ChatGPT enterprise users can save an average of 40 to 60 minutes of work time per day, while highly skilled users can save more than 10 hours per week. GPT-5.2 is considered a key version for further amplifying productivity.
Three versions were released simultaneously to fully cater to different professional usage scenarios.
OpenAI stated that GPT-5.2 is being released in three versions: Instant, Thinking, and Pro, focusing on daily efficiency, deep inference, and the quality of high-difficulty problems, respectively. In the ChatGPT application, GPT-5.2 will initially be available to users with paid plans. The API platform is fully open to developers.
Significant leap forward in professional competence, reaching the level of human experts at GDPval for the first time.
OpenAI points out that GPT-5.2 Thinking represents a key breakthrough in the GDPval professional knowledge job assessment. This assessment covers 44 occupations across 9 industries and requires models to directly produce deliverables such as presentations, spreadsheets, schedules, and business documents.
The results showed that GPT-5.2 Thinking outperformed or matched industry experts in 70.9% of the tasks, with a production speed 11 times faster than human experts and a cost of less than 1%, making it the first OpenAI model to reach or even surpass human professional levels.
Practical skills have been comprehensively upgraded, and project output and program development have been strengthened simultaneously.
In practical applications, GPT-5.2 Thinking offers more complete presentation and spreadsheet structures with logic closer to that of professional consultants and analysts, and can handle complex tasks such as human resource planning, equity structure, and project management. In investment banking-level spreadsheet tasks, its performance is approximately 9% better than the previous generation.
In software development, GPT-5.2 Thinking achieved an accuracy of 56% in the SWE-Bench Pro real-world code repair test. It can directly read the code repository, understand engineering problems, and output usable patches.
(Note: SWE-Bench Pro is a test that simulates "real-world software engineering work," requiring AI to directly fix programming problems in actual projects, rather than just answering theoretical questions.)
With improved credibility and integration, collaboration on long files, images, and tools has become more mature.
OpenAI states that GPT-5.2 Thinking's response error rate is approximately 30% lower than its predecessor. In the MRCRv2 benchmark, it maintains near-perfect accuracy even with extremely long documents of hundreds of thousands of words, making it suitable for contract, financial statement, and verbatim transcript analysis.
In terms of image understanding, GPT-5.2 significantly reduced the error rate in interpreting charts and software interfaces. In terms of tool calling capabilities, it achieved an accuracy of 98.7% in the Tau2-bench test and can stably complete the entire task process across systems and multiple steps.
(Note: MRCRv2 is a test specifically designed to assess an AI's ability to accurately grasp key information within extremely long content and avoid confusing the context during multiple inferences. Tau2-bench is an evaluation tool that simulates real customer service and business processes, used to test whether an AI can correctly call tools, integrate information, and complete the entire task in multiple rounds of interaction.)
This article, "OpenAI's New GPT-5.2: Capable of Creating Mini-Games and Greeting Cards, Saving Enterprise Users 10 Hours of Work Time Per Week," first appeared on ABMedia .
