Just now, OpenAI officially released GPT-5.5 Instant, making it the default model for ChatGPT, replacing the previous GPT-5.3 Instant, and making it available to all users.
The Instant series is ChatGPT's main daily model, used by hundreds of millions of users every day. The official statement says that at this scale, even small improvements accumulate to a considerable effect. This version focuses on three things: greater accuracy, greater simplicity, and a better understanding of your needs.
Compared to the previous version, the new model maintains low latency while significantly improving accuracy, response style, and personalization capabilities.
The improvement in accuracy is most pronounced in high-risk areas. Internal testing shows that GPT-5.5 Instant has reduced the illusion rate for medical, legal, and financial questions by 52.5% compared to the previous version. The error rate for previously flagged incorrect conversations by users has also decreased by 37.3%.
In addition to text-based question and answer, the ability to analyze images and photos, the quality of answers to science questions, and the ability to determine when to actively use search tools have all improved.
The improvement in math and science skills was even greater. In the AIME 2025 math competition, the GPT-5.5 Instant score was 81.2, while the GPT-5.3 Instant score was only 65.4.
The scores on the Doctoral Science Test GPQA rose from 78.5 to 85.6, the Multimodal Reasoning Benchmark MMMU-Pro rose from 69.2 to 76, the Scientific Graph Comprehension CharXiv score rose from 75 to 81.6, and the document parsing error rate fell from 14.6% to 12.5%.
OpenAI demonstrated the difference between the two versions using an algebra problem. A user submitted a solution to a radical equation and asked if it was correct. GPT-5.3 Instant found that substituting x=3 into the original equation did not work, and directly determined "no real solution," without further investigation. GPT-5.5 Instant also found that x=3 was invalid, but then located the specific error in the user's expansion of (x-1)² and provided the correct solution.
The response style is also a key focus of this update. The new model is shorter, no longer piling on formatting and emojis, and reduces unnecessary follow-up questions. The official example is a common scenario: asking how to tactfully get a talkative colleague to talk less.
GPT-5.3 Instant provides five categorization strategies and includes a "don't do" list, which is well-structured but slightly excessive. GPT-5.5 Instant's responses are 30.2% shorter and 29.2% shorter, with a tone more like a friend's advice, focusing on how to steer the conversation toward one's own needs rather than the other person's speaking habits.
Personalization is another key feature of this update . Plus and Pro users can have the model access historical conversations, uploaded files, and associated Gmail content to get more personalized responses without having to re-explain the context each time.
The official documentation presents a comparison of teahouse recommendations: GPT-5.3 Instant only knew the user was in San Francisco and recommended a few popular general shops. GPT-5.5 Instant, however, found records from the user's conversation history indicating frequent visits to Asha Tea House and a preference for high-mountain tea over heavily sweet milk tea, and based on this, recommended Ceré Tea and Song Tea & Ceramics, which are more suitable in style, and explained the reasons for the recommendations.
At the same time, the "Memory sources" feature will be available on all consumer versions. When personal background information is used in an answer, users can see which historical conversations or saved memory entries were called, and can delete or correct outdated content at any time.
For example, when a user asks for dinner suggestions for the week, ChatGPT recommends a miso salmon bowl based on memories such as "preparing for a marathon," "preferring a light, high-protein diet," and "liking cookies," and lists the memory sources used in this answer in the Sources panel on the right. Users can also mark a single memory as relevant or irrelevant, make corrections, view all memories, or delete the memory directly.
OpenAI states that this view displays the most relevant sources and may not cover all records retrieved by the model; it will continue to be improved. Users who do not wish to be recorded can also choose a temporary conversation mode, which does not read or update any memories. When sharing a conversation, the other party will not see these source records.
GPT-5.3 Instant will remain available to paid users for three months before being officially discontinued. Personalization features are currently available on the web version for Plus and Pro users, with rollout to mobile, free, Go, and enterprise versions planned for the coming weeks, varying by region.
For developers, GPT-5.5 Instant is available via API under the name "chat-latest".
Oh, by the way, OpenAI is also hosting an AI-driven party today. In a Stripe Sessions conversation, Altman mentioned that while preparing for the GPT-5.5 launch party, he casually asked the model: "What kind of party do you want?" The model seriously provided a list. It wanted the party to be on May 5th (US time), with as few speeches as possible, and a human creator to give a toast, but it itself didn't want to go on stage to toast.
It also proposed setting up a dedicated session to collect GPT-5.6 suggestions and feed them back to the model. Altman said these requests were "wonderful" and would ensure the party ran smoothly. The time was ultimately set for 5:55 p.m., also the model's own choice. The party location was chosen to be at OpenAI's San Francisco headquarters, with OpenAI covering the airfare and hotel expenses for non-local guests.
The list of invitees was selected by Codex from the tweet replies, and the registration link closed at 5:55 PM on April 30th. Over 8,000 people registered within 24 hours, and some users have already shared photos of the invitation emails they received. Those who weren't selected also received an email informing them that OpenAI had increased their Codex access quota tenfold.
Ultraman also responded to users' jokes: Musk can come if he wants; the world needs more love. That being said, unfortunately, Musk's love is currently all in the lawsuit against OpenAI, so the champagne celebrating GPT-5.5 can only be enjoyed by Ultraman himself.
Here is the link to the OpenAI blog:
https://openai.com/index/gpt-5-5-instant/
This article is from the WeChat official account "APPSO" , authored by Discover Tomorrow's Products, and published with authorization from 36Kr.
