OpenAI introduced a new version of the chatbot ChatGPT based on the GPT-5.2 model. The company calls it the most advanced AI for professional work and complex tasks. However, user reactions have been mixed: shortly after the release, the model faced a wave of criticism.
BeInCrypto's editorial found out what's wrong with GPT-5.2 and why part of the audience remains dissatisfied with the update.
The best Russian neural networks in 2025: top 8 services for text, images, and business
What is GPT-5.2
GPT-5.2 is the new flagship model of OpenAI, primarily focused on work tasks rather than everyday communication. Developers position it as a tool for specialists who need to solve complex tasks and bring them to a finished result.
According to OpenAI, GPT-5.2 has been improved in several areas:
creating tables, presentations, and working documents;
writing, analyzing, and refining code;
working with images, graphs, and interfaces;
understanding long texts and large documents;
performing multi-step tasks using built-in tools.
The company claims that the model has become more useful precisely where it is important not just to give an answer but to complete a full working task.
Test results and performance
OpenAI provides data from internal and public tests that should confirm the model's progress:
In the GDPval test, GPT-5.2 outperformed or matched professionals in 70.9% of tasks across 44 professions;
the model creates working materials faster and cheaper than specialists, provided there is subsequent verification by a human;
the quality of tables, presentations, and reports has become noticeably higher compared to the previous version.
According to experts, GPT-5.2 performs better on applied tasks where structure, formatting, and logic of the result are important.
In the field of development, OpenAI highlights the following changes:
the best results on SWE-Bench Pro and SWE-bench Verified tests;
more stable performance with large codebases;
improved error correction and refactoring;
enhanced capabilities in frontend development and complex interfaces.
Early users note that the model has become more convenient for everyday work of programmers and requires fewer manual edits.
Accuracy and working with long texts
Another area of improvement is the reliability of responses and context:
GPT-5.2 makes factual errors less often compared to GPT-5.1;
the model retains logic better when working with long documents;
increased accuracy when analyzing reports, contracts, and multi-page files.
OpenAI emphasizes that the model still requires verification in critical tasks, but overall it has become more stable.
Working with images
GPT-5.2 has also received improvements in visual tasks:
more accurate understanding of graphs and diagrams;
better interpretation of interfaces and screenshots;
correct working with technical diagrams and visual reports.
Such capabilities are aimed at specialists in finance, engineering, design, and support services.
Availability and model versions
In ChatGPT, GPT-5.2 is available in three versions:
Instant is the fast version for everyday tasks;
Thinking is for complex tasks and detailed analysis;
Pro is the most accurate option for critical requests.
The model is gradually becoming available to paid users. The API for GPT-5.2 is more expensive than the previous version, but OpenAI claims that due to quality and time savings, the overall efficiency is higher.
Criticism of GPT-5.2
Users found that GPT-5.2 avoids direct answers to simple logical tasks. It provides a formally safe explanation, whereas the older AI model immediately understands the essence of the question and provides a correct example.
One user published evidence online:
He asked the classic philosophical question about the paradox—a statement that is both true and false.
ChatGPT-4o immediately provided the canonical example—'This sentence is false', that is, the liar's paradox, and correctly explained why it is considered both true and false.
GPT-5.2 refused to present the paradox at all and resorted to an academic explanation about 'classical logic' and context, effectively bypassing the essence of the question.
Because of this, users concluded that GPT-5.2:
less well understands the meaning of non-standard or philosophical questions;
more often resorts to safe, template formulations instead of direct answers;
seems 'dumber' on simple intellectual tests where previous models demonstrated flexibility of thought;
excessively formal and cautious even where there is no risk.
The criticism is not that GPT-5.2's answer is logically incorrect, but that it is weaker in the quality of thought: the model did not recognize the obvious paradox and instead of demonstrating intelligence, it limited itself to refusal. This is what users interpreted as a step back and a sign of degradation of 'smart' responses.
There are also much more obvious problems. Users are widely complaining that GPT-5.2 is unable to answer even some of the simplest questions. For example, the screenshot below shows that the AI could not determine how many letters 'r' are in the word strawberry.
And the screenshot below shows that GPT-5.2 cannot determine the number of letters 'r' in the word garlic.
Many users online contrast the new version of the Gemini neural network with GPT-5.2. In their opinion, the latter demonstrates far more outstanding results.
"Users are trying GPT-5.2 after several weeks of using Gemini 3."
Want to gain access to expert insights? Subscribe to our Telegram channel, get access to trading signals and market news, communicate with our analyst. Stay one step ahead of the market every day!
