New AI OpenAI GPT-5.2 outperforms human experts in 70% of tasks: is the end of humanity near?

OpenAI announced the launch of the GPT-5.2 line, marking it as the most noticeable update in the direction of universal AI. The new generation of models has improved work with code, expanded the scope of context and strengthened agent functions. The models are presented in the Instant, Thinking and Pro versions, and are already starting to appear for users of ChatGPT Plus, Enterprise, Business, as well as API.

The main focus of the release was the results in GDPval – a benchmark that assesses the ability to perform clearly described work tasks in several dozen specialties. According to OpenAI, the GPT-5.2 Thinking model for the first time demonstrated indicators comparable to the level of experts: in blind tests it surpassed or repeated the results of professionals in more than 70% of tasks. Test cases included the preparation of complex tables, creating presentations and working with technical documentation. At the same time, the model performed the tasks significantly faster than a person and with minimal costs.

AI Benchmarks

An important indicator for developers was the score in SWE-Bench Pro, where GPT-5.2 Thinking achieved a result of 55.6%, working with real software engineering tasks and multilingual code bases.

The model also received improved visual capabilities. OpenAI notes that the number of errors in interpreting graphs, diagrams and interface layouts has been reduced by almost half. The number of hallucinations has also been reduced by about a third compared to the previous version. Internal tests with MRCRv2 showed almost one hundred percent accuracy when analyzing documents with a volume of up to 256 thousand tokens, which is oriented to industries working with large amounts of information.

It is known that companies such as Notion, Zoom, Shopify and Databricks have already implemented the new model and report significant performance gains in agent scenarios, where AI independently uses tools to perform complex tasks.

Despite the launch of the update, OpenAI is not yet disabling the GPT-5.1 models – they will remain available for about three more months.