OpenAI's new GPT-5.2: Can be used for mini-games and greeting cards, saving enterprise users 10 hours of work time weekly

OpenAI launched GPT-5.2 on 12/12, positioning it as the most suitable AI model series for “professional knowledge work” currently. Officially, GPT-5.2 can directly produce practical work results such as presentations, spreadsheets, code, and long document analysis, supports cross-tool, multi-step task processing, and can also create front-end web mini-games and greeting cards. According to feedback, ChatGPT enterprise users save an average of 40 to 60 minutes of work per day, with highly engaged users saving over 10 hours per week. GPT-5.2 is regarded as a key version to further amplify productivity.

Three versions launched simultaneously, fully supporting different professional scenarios

OpenAI states that GPT-5.2 is released in three versions: Instant, Thinking, and Pro, each focusing on daily efficiency, in-depth reasoning, and high-difficulty problem quality, respectively. For ChatGPT applications, GPT-5.2 will be available first to paid plan users. On the API platform, it is now fully open for developers to use.

Significant leap in professional work capabilities, first reaching human expert level in GDPval

OpenAI points out that GPT-5.2 Thinking has made a breakthrough in the GDPval professional knowledge work assessment. The evaluation covers 44 occupations and 9 major industries, requiring the model to produce deliverables such as presentations, spreadsheets, schedules, and business documents directly.

Results show that 70.9% of tasks with GPT-5.2 Thinking are judged to surpass or match industry experts, with output speed exceeding human experts by 11 times and costs below 1%. This marks OpenAI’s first model to reach, and even surpass, human professional standards.

Comprehensive upgrade in practical capabilities, enhanced project output and coding development

In practical applications, GPT-5.2 Thinking produces more complete presentation and spreadsheet structures with logic closer to professional consultants and analysts. It can handle complex content such as workforce planning, equity structures, and project management. In investment banking-level spreadsheet tasks, its performance improves by approximately 9% compared to the previous generation.

In software development, GPT-5.2 Thinking achieves a 56% accuracy rate in the SWE-Bench Pro real-world code repair test, capable of directly reading code repositories, understanding engineering problems, and outputting usable patches.

(Note: SWE-Bench Pro is a set of tests simulating “real-world software engineering work,” requiring AI to directly fix actual project code issues rather than just answer theoretical questions.)

Improved credibility and integration, more mature handling of long documents, images, and tool collaboration

OpenAI states that GPT-5.2 Thinking’s error rate is approximately 30% lower than the previous generation. In the MRCRv2 evaluation, even with ultra-long documents of tens of thousands of words, it maintains near-perfect accuracy, suitable for contract, financial report, and transcript analysis.

In image understanding, GPT-5.2 significantly reduces errors in interpreting charts and software interfaces. In tool invocation capability, it achieves an accuracy of 98.7% in Tau2-bench testing, capable of reliably completing cross-system, multi-step full task workflows.

GPT-5.2 can also create front-end web pages, mini-games, and more.

(Note: MRCRv2 is a test designed to evaluate AI’s ability to accurately grasp key information within ultra-long content and maintain context across multiple reasoning steps. Tau2-bench simulates real customer service and business workflows to test whether AI can correctly call tools, integrate information, and complete entire tasks in multi-round interactions.)

This article “OpenAI launches GPT-5.2: capable of making mini-games and greeting cards, saving enterprise users 10 hours of work weekly” first appeared on Chain News ABMedia.