OpenAI анонсировала новейшие ИИ-модели GPT-5.2, устанавливающие новый стандарт в производительности Translated: OpenAI announces the latest AI models GPT-5.2, setting a new standard in performance.

OpenAI has unveiled its range of GPT-5.2 models. As of December 12, the Instant, Thinking, and Pro versions are available to all users.

«Currently, ChatGPT Enterprise users save an average of 40-60 minutes daily thanks to AI, with active users saving over 10 hours a week. We designed GPT-5.2 to unlock even greater economic potential for individuals,» the startup stated in its blog.

The Thinking model has shown remarkable results in several assessments.

It achieved expert-level performance in the GDPval benchmark, which evaluates intellectual task execution in 44 professions. Test tasks include creating presentations and working with spreadsheets.

The «Thinking» version of the large language model (LLM) completes GDPval benchmark tasks 11 times faster than human experts, with operational costs amounting to less than 1% of a specialist’s expenses.

The company emphasized that GPT-5.2 Thinking «raises the bar for professional work.» It showcases:

The GPT-5.2 Instant version is designed for everyday work and learning. It has a warm and conversational style, clear explanations highlighting key information, improved step-by-step guides, and quality translations of technical information.

GPT-5.2 Pro is promoted as the most powerful solution for complex queries. The neural network demonstrates high efficiency in specialized fields, including programming and scientific research.

“GPT-5.2 is part of an ongoing process of model enhancement. We continue to address known issues, such as unwarranted refusals and delays, to make the product more beneficial,” OpenAI emphasized.

GPT-5.1 will remain accessible to paid users for three months.

GPT-5.2 Thinking has set a new record in the SWE-Bench Pro test with a score of 55.6%. This test evaluates the model’s ability to operate across four languages.

A high score of 80% was also achieved in the SWE-bench Verified test.

“For everyday professional use, this means that the model can debug code more reliably, implement requests for new features, refactor large codebases, and make end-to-end corrections with minimal manual intervention,” OpenAI’s blog stated.

GPT-5.2 Thinking performs better in frontend development and creating complex, non-standard interfaces compared to GPT-5.1 Thinking.

“GPT-5.2 represents the most significant advance for GPT models in the domain of agent programming since GPT-5 and is the best solution in its price range,” OpenAI noted.

GPT-5.2 Thinking experiences fewer hallucinations compared to GPT-5.1 Thinking. The model proves more reliable in daily information handling, conducting research, writing text, analysis, and decision support.

GPT-5.2 Thinking «sets a new standard» in reasoning with long context. It achieved leading results in the OpenAI MRCRv2 test, which assesses the model’s capability to integrate information distributed across lengthy documents.

In real-world deep analysis tasks requiring coherent information over hundreds of thousands of tokens, GPT-5.2 Thinking is «significantly more accurate» than GPT-5.1 Thinking.

GPT-5.2 Thinking is OpenAI’s most powerful visual perception model. It reduces the number of errors in analyzing charts and understanding interfaces by about half.

The neural network can interpret dashboards, screenshots, technical graphs, and reports more accurately.

As demonstrated, GPT-5.2 successfully identifies key areas and defines object boundaries. In contrast, GPT-5.1 only highlights individual fragments, showing a weak grasp of spatial structure.

Although both models can make mistakes, version 5.2 excels at image analysis significantly better.

It’s important to note that OpenAI planned the release of GPT-5.2 in December as a response to the rising popularity of Google’s Gemini.