The End of Made-Up Facts? OpenAI Measures Progress in Numbers
The main pain point of generative artificial intelligence is hallucinations – situations where the model confidently presents information that is not true. OpenAI now claims that with its latest model GPT-5.5 Instant, this problem has significantly eased. According to the company's internal evaluations, the model generated 52.5% fewer hallucinated claims than GPT-5.3 Instant in so-called high-stakes areas such as medicine, law, or finance. In conversations that users themselves marked as factually incorrect, inaccurate statements dropped by 37.3%.
The company published the data on May 5, 2026, while simultaneously beginning the gradual rollout of the new model to all ChatGPT users. For paid users, the older GPT-5.3 Instant will remain available in settings for another three months before being permanently retired. This approach reflects past reactions from the community, which often grew attached to departing models and expressed almost nostalgic fondness for them.
What Changed in Practice
GPT-5.5 Instant brings not only better factual accuracy. OpenAI promises that the new model is generally more capable in everyday tasks. It better analyzes uploaded photos and images, answers questions from STEM fields (science, technology, engineering, mathematics), and more intelligently decides when it's appropriate to search for current information on the web instead of relying on training data.
A significant change is also the communication style. Responses are to be more concise, direct, and less over-formatted. The model should avoid unnecessary follow-up questions and excessive use of emojis, which should improve readability especially for more professional queries. Compared to its predecessor, GPT-5.5 Instant uses more than 30% fewer words in some scenarios without losing content.
Visible Memory and Greater Personalization
Another new feature is better use of context. GPT-5.5 Instant draws more effectively from previous conversations, uploaded files, or a connected Gmail account so that responses are more personal and relevant. This means that if a user previously discussed specific preferences or projects, the model remembers them better and builds upon them.
However, the crucial new feature is „memory sources“, which is being introduced across all ChatGPT models. When the model personalizes a response based on earlier interactions or stored memories, the user will now see which specific context was used. They can thus easily delete outdated information, correct inaccuracies, or remove sensitive data. This feature should increase trust in the tool, especially when working with personal data. However, note that if you share a chat with someone else, these sources will not be visible to them.
However, the availability of personalization depends on the region, so users in the Czech Republic should expect that some features may initially be more limited than in the USA. Enhanced personalization is first being made available to ChatGPT Plus and Pro subscribers on the web, with the mobile app and free users coming in the following weeks.
What This Means for Czech Users and Businesses
ChatGPT is fully available to Czech users including Czech language support, albeit with minor limitations compared to English. The arrival of GPT-5.5 Instant thus brings concrete benefits to Czech individuals and businesses: faster and more accurate responses, better processing of visual materials, and greater transparency when working with personal data.
However, for businesses in the Czech Republic and the European Union, it is important to maintain a cautious approach. Although a drop in hallucinations of more than half sounds significant, in regulated industries – such as healthcare, law, or financial advice – EU AI Act rules apply requiring human oversight over automated decisions. Czech companies should therefore use GPT-5.5 Instant as an assistant, not as a replacement for an expert. The new memory sources feature also helps meet GDPR requirements for transparency in personal data processing, because the user has an overview of what data the model uses.
The Future of Factual Accuracy in Language Models
While OpenAI emphasizes reducing hallucinations, the competition is not standing still. Google is intensively investing in the personalization of its Gemini model, while Anthropic focuses on the safety and reliability of Claude. A 52.5% reduction in error rate is a significant step, but it does not mean that hallucinations in large language models are solved. Users should still verify key information, especially if they use it for important decisions.
GPT-5.5 Instant is available for free to all ChatGPT users, while paid Plus ($20 per month) and Pro ($200 per month) plans offer access to more powerful model variants and priority processing. For developers, the model will be available in the API as chat-latest.
Can I turn off the new GPT-5.5 Instant model in ChatGPT and return to the older version?
If you have a paid Plus or Pro account, you can select the older GPT-5.3 Instant in ChatGPT settings for another three months before it is permanently retired. Free version users cannot avoid the new model.
How does the „memory sources“ feature work and is it safe for sensitive data?
Memory sources show what context from your previous conversations or stored memories ChatGPT used to personalize the response. You can delete or edit this context at any time. If you share a chat, memory sources will not be displayed to the other party. However, you should still handle sensitive personal data carefully, especially when working in compliance with GDPR.
Is GPT-5.5 Instant better than Gemini or Claude at fighting hallucinations?
OpenAI has only published an internal comparison with its own previous model GPT-5.3 Instant. A direct independent comparison with Gemini 3.1 Pro or Claude Opus 4.7 in the area of factual accuracy is not yet publicly available. Users should always verify key facts regardless of the model used.
Title: ChatGPT Hallucinates Half as Much. OpenAI Launches GPT-5.5 Instant for All Users
Perex: The new default ChatGPT model promises 52.5% fewer hallucinations in medicine, law, and finance. It brings more concise responses, better personalization, and a new visible memory feature for greater control over data.
