OpenAI dumps GPT-5.5 Instant! Illusion plummets 52%, talks 30% less, all free!

Today.OpenAIOfficially launchedGPT-5.5 Instantversion, will be rolled out gradually to all ChatGPT users starting today, replacing GPT-5.3 Instant as the default model.

The update focuses on daily interactions, with GPT-5.5 Instant delivering a more natural conversational tone, more accurate and compact answers, and the model's ability to pull up past conversations to supplement contextual information when the user uses the personalization feature.

Sam Altman was the first to retweet the official announcement tweet “pushing” the model, stating, “The combination of improvements in speed, intelligence, and individuality, coupled with a strong ability to remember and personalize, when all working together, gives an experience that is much more than the simple addition of the parts. simply added together, but rather an experience where the whole is greater than the sum of its parts.”

In internal evaluations, in the fields of medicine, law, and finance, GPT-5.5 Instant had a reduced rate of hallucinations compared to GPT-5.3 Instant52.5%.

For benchmarking, in CharXiv-reasoning, which measures the accuracy of scientific diagrammatic reasoning, GPT-5.5 Instant improved over GPT-5.3 Instant by6.6%. Accuracy improvement of GPT-5.5 Instant in the multimodal expert reasoning test MMMU-Pro6.8%.

In the document parsing task, GPT-5.5 Instant has reduced the error rate of the2.1%The relative decline is about14.4%.. On the PhD-level science quiz test, the GPT-5.5 Instant improved accuracy by7.1%.. In the math competition AIME 2025, its accuracy went up by15.8%.

In the API, GPT-5.5 Instant is named “chat-latest”. For paid subscribers, GPT-5.3 Instant will remain available for three months before it is retired and can be accessed through the model configuration settings.

Enhanced personalization based on past conversations, uploaded files and connected Gmail is being rolled out to Plus and Pro users on the web side, coming soon to mobile, with plans to expand to Free, Go, Business and Enterprise users in the coming weeks.

The Memory Sources feature is rolling out to all ChatGPT Personalized Package subscribers on the web side and will be coming to mobile soon. Availability of specific personalized sources may vary by region.

Below OpenAI's official announcement tweet, users noted the improvement in the model's AIME score, arguing that “this is ostensibly a product update, but it's a pure reasoning upgrade, not just a chatty tweak. It's a “sneaky” way to release a thinking model.”

Other users found that, “The ‘warmer and cleaner‘ points are exactly what users have really complained about. Interestingly, the biggest model upgrade of the year is essentially more of a ’character patch”."

However, there are many netizens who are not sold on this upgrade, they want more useful feature updates. There are even netizens who are missing the GPT-4o.

I. Improvement in image parsing ability and reduction of false information by 52.5%

In internal evaluations, for high-risk tips covering the medical, legal, and financial fields, GPT-5.5 Instant generated fewer false information than GPT-5.3 Instant did52.5%. In particularly challenging conversations where users had flagged the presence of factual errors, it also reduced the number of37.3%of inaccurate statements.

GPT-5.5 Instant improves image parsing, quizzing in STEM subjects (science, technology, engineering, and math), and intelligently determines whether or not to invoke a web search to give a better quality response.

As you can see from the case study, GPT-5.5 Instant initially recognized the incorrect solution, but then realized that it did not hold when substituting x=3 back into the original equation. It recognized the actual algebraic error (the user shifted terms incorrectly) and then used the rooting formula to arrive at the correct solution.

The GPT-5.3 Instant, on the other hand, while also finding that x = 3 does not hold, stops there and incorrectly concludes that there is no real number solution rather than rechecking the algebraic steps and solving the corrected quadratic equation.

II. More compact answers and 30.21 TP4T fewer words

In addition, GPT-5.5 Instant's answers are more compact and to the point, while maintaining a sense of warmth and personalization.

The model reduces lengthy and overly formatted answers to questions that are too long while conveying the same information and being more useful. It also reduces unnecessary follow-up questions and avoids cluttering responses with emoticons, for example.

GPT-5.5 Instant uses a reduced number of words30.2%The number of lines has been reduced29.2%The tone of the responses was appropriate: informal, practical and appropriate to the workplace, while avoiding over-explanation. The tone of the responses is appropriate: informal, practical and workplace-appropriate, while avoiding over-explanation. The model provides scripts that can be practically used in different situations, always framing the issue around “boundaries”.

The responses in GPT-5.3 were more complete, especially the “what not to do” section, but were a bit too complex for an informal, daily advice type of prompt, with a structure and level of sophistication that may have exceeded the user's actual needs.

Third, automatically retrieve the history of the dialog, memory source function of the whole system on-line

GPT-5.5 Instant also utilizes past chats, uploaded files, and contextual information from Gmail to personalize responses.

The model intelligently determines when to incorporate personalization elements to optimize responses, while its speed of retrieving historical conversations and matching context is dramatically improved, eliminating the need for users to repeat expressions over and over again.

As can be seen, GPT-5.5 Instant responses are better able to cite past conversations and relevant connected data to provide more nuanced and highly personalized recommendations. The GPT-5.3 Instant responses, on the other hand, while taking into account the factor that the user is located in San Francisco, still give more generalized suggestions for places to recommend trying.

ChatGPT's full range of models is now live with the Memory Sources feature. Users can view the contextual basis cited for personalized responses and gain autonomous control.

When the model generates a personalized response, the user is able to trace the specific context in which the answer was given, including saved memories and the history of the conversation; outdated, invalid or irrelevant information can be deleted and corrected at any time.

In addition, when users share the content of the conversation, the source information of the memory will not be shown to the public. At the same time, this feature supports various privacy control methods: you can individually delete historical conversations that you don't want to be quoted, edit or clear the saved memories in the setting interface, or use the temporary conversation mode, which doesn't recall or update personal memories throughout the whole process.

Conclusion: Interaction Quality and User Controllability Improvement

In the context of convergence of underlying capabilities, “How can we make models answer more comfortably to users?” becomes a question for big model vendors to think about.

The GPT-5.5 Instant update gives OpenAI's answers:: one, its reduction in the rate of illusions in expertise quizzing; two, conciseness of answers and tone modulation are incorporated into the optimization goals; and three, the memory source function establishes a foundation of trust.

Objectively speaking, it is difficult to fully quantify the value of this type of “experiential update” through traditional benchmarking, and its real effect will depend on the subjective experience of users in long-term use.