Finally, GPT-5.5 Instant is here too! Earlier, OpenAI Officially launch GPT-5.5 Instant Modelat the same time, also ChatGPT The default model has been instantly switched from GPT-5.3 Instant to GPT-5.5 Instant, meaning free tier users can also immediately experience this model. The improvements this time come in three major areas: “More accurate answers,” “Shorter responses,” and “Better understanding of you.” Here’s a breakdown for everyone.
GPT-5.5 Instant Launches: Major ChatGPT Default Model Upgrade Cuts Hallucinations by 52.5%, Delivers Conciser and More Personalized Responses
First is accuracy—the hallucination rate is cut in half, which simply means improved factual accuracy.
OpenAI says that in internal testing comparing GPT-5.5 Instant with the previous generation GPT-5.3 Instant, on high-stakes questions in medical, legal, and financial domains where errors are absolutely unacceptable, GPT-5.5 Instant had 52.5% fewer hallucinated claims compared to the previous generation. Additionally, among conversation samples where users actually flagged “you got this wrong,” erroneous claims also decreased by 37.3%.
For specific improvements, refer to the benchmark tests: GPT-5.5 Instant scored 81.2% on the AIME 2025 math exam, compared to 65.4% for the previous generation GPT-5.3 Instant; multimodal reasoning MMMU-Pro also improved from 69.2% to 76%:
Besides this, there have also been improvements in image analysis, STEM Q&A, and its ability to judge when to open a search tool on its own.
According to news reports from OpenAI, when GPT-5.3 Instant found that substituting x=3 back into the original equation didn’t work, it immediately concluded “no solution.” GPT-5.5 Instant, on the other hand, would trace back through the calculation to find where the error occurred, ultimately discovering that the user had missed the -2x term when expanding (x-1)².
In terms of answer length, using the topic “how to tell a colleague they talk too much” for comparison, GPT-5.5 Instant’s response contains 30.2% fewer characters and 29.2% fewer lines than GPT-5.3 Instant.
The difference between the two responses is that GPT-5.3 would list “1. Polite deflection, 2. Gentle boundary-setting, 3. Immediate interruption, 4. Preventive defense, 5. Slightly firm tone” along with a “What not to do” list, complete with markdown headings.
GPT-5.5 directly provides five phrases of varying directness, focusing on “you’re busy and need to focus” rather than “they talk too much,” and adds a note: “Colleagues who talk a lot are generally not malicious, it’s just habit. A clean redirect usually works better than dramatic conflict.”
Same amount of information, half the space. OpenAI deliberately addressed issues like responses being too long, excessive formatting, unnecessary follow-ups, and random emoji.
The more personalized experience that truly understands you is the biggest upgrade this time around. GPT-5.5 Instant will more proactively use your past chat history, uploaded files, connected Gmail (if you have it enabled), and more as context for its responses.
For example, OpenAI showed that GPT-5.3 Instant gave a recommendation list for the SF area (Heytea, Junbi Matcha, Aroma Tea Shop, etc.). Although it did pick up on the key information that you’re in SF, it still gave tourist-level recommendations.
GPT-5.5 Instant noticed the user often visits Asha Tea House, prefers Taiwanese high mountain tea, and doesn’t like overly sweet bubble milk tea, so it recommended Ceré Tea and Song Tea & Ceramics. It even helped make the decision: “If you’re going tomorrow, go to Ceré; if you want something more special, choose Song Tea.”
The default model for ChatGPT has now been changed to GPT-5.5 Instant, but if you really like GPT-5.3 Instant and are a Plus or Pro paid subscriber, you can still switch back from the model settings within three months before the old model is officially retired.
Additionally, personalized features will initially roll out for Plus and Pro users on the web version, with the mobile version coming later. Free, Go, Business, and Enterprise users can expect a gradual rollout over the next few weeks.
OpenAI has also rolled out a new “Memory sources” feature, so every time ChatGPT gives a personalized response, you can see which past conversations and stored memories it drew from, and you can instantly delete or correct any outdated information.
Source: KOCPC Chinese