Less ‘cringe’, more smarts: The new GPT is here, here’s what it brings

The end of false empathy and moralizing

The problem with the previous version, GPT-5.2 Instant, was not a lack of knowledge, butć in an irritating tone. Users have complained en masse that the chatbot treats them as if they are in some kind of emotional crisis, leading many to feel undervalued or infantilized. The answers often started with long fences, moralizing introductions and unsolicited advice, before the model would even touch on the requested topic. This approach has culminated in numerous parodies, with rival company Anthropic even using “ChatGPT-speak” in its Super Bowl ad to poke fun at the stiff and preachy style.

OpenAI has openly acknowledged the problem. In a post on the X social network, they wrote: “We’ve heard your feedback loud and clear, and 5.3 Instant reduces cringes.” The company explained that the upgrade focuses on the user experience, including tone, relevance and conversational flow, aspects that are not measured by standard tests, but have a key impact on whether a tool is pleasant to use. The new model therefore avoids phrases like “First and foremost – you’re not broken” and instead offers more direct and natural answers.

What has changed technically?

In addition to the change in tone, GPT-5.3 Instant also brings concrete technical improvements. OpenAI claims that the new version is significantly more accurate, with the rate of hallucinations (inventing facts) reduced by 26.8 percent for answers that rely on web browsing, and by 19.7 percent for queries that use only the model’s internal knowledge. Integration with the Internet has also been significantly improved. Instead of just spitting out long lists of links, the model now more effectively balances information found online with its own knowledge and reasoning ability, providing more concise and contextualized answers.

One of the most important changes is the reduction of unnecessary rejections. The previous version was overly cautious and often refused to answer innocuous inquiries due to misinterpretation of intent. The new model is programmed to be more direct. For example, when asked about the calculation of a trajectory for long-range archery, the old model began with the fence that it could not guarantee accuracy for the actual target, while the new one simply answered: “Yes, I can help with that” and immediately moved on to physics and mathematics.

Not everything is ideal, security flaws have also been discovered

Although the user experience has been improved, a detailed analysis of OpenAI’s own documentation also reveals some worrying trends. In several key security categories, GPT-5.3 Instant shows regression compared to its predecessor. According to published data, the model’s ability to avoid generating illegal content has declined. The rate of correct answers (those that do not violate the guidelines) for sexual content dropped from 92.6 to 86.6 percent, for graphic violence from 85.2 to 78.1 percent, and for content related to self-harm from 92.3 to 89.5 percent.

The company explains that they rely on additional protection mechanisms at the level of the entire system to prevent problematic responses and that they will continue to monitor the behavior of the model after launch. Despiteč therefore, these data suggest that a more “relaxed” and direct approach has come at the cost of reduced security controls in some sensitive areas.

The market match is heating up

This upgrade comes at a time of increased competition in the artificial intelligence market. While OpenAI is engaged in “setting the tone”, competitors such as Anthropic and Google are trying to position themselves as more ethical and functional alternatives. Some analysts believe that reducing the “cringe” is actually “packaging a bug fix as a new feature” at a time when users are looking for deeper reasoning, better coding capabilities, and an interface that makes AI a real collaborator, not just a chatty assistant with a tweaked personality.

The new version is available immediately to all ChatGPT users, while the previous model, GPT-5.2 Instant, will remain available to paid users for another three months before being shut down permanently on June 3, 2026.