GPT-5.5 Instant is officially live today, and all ChatGPT users can use it for free!
It comes with three major highlights: more concise answers, stronger memory, and better personalization.
GPT-5.3 Instant’s hallucination rate has been halved, and its score on the AIME math competition has jumped from 65.4% to 81.2%.
Optimized math calculation model
Internal evaluation shows that for prompts involving high-risk fields such as medicine, law and finance, the hallucinations generated by GPT-5.5 Instant are 52.5% less than those of GPT-5.3 Instant.
When the previous generation, GPT-5.3 Instant, was released, the improvement in hallucinations was about a little over 20%, but this time it has more than doubled. In the difficult conversations that users actively marked as “factual errors”, inaccurate statements have also decreased by 37.3%.
For example, in the actual test below, GPT-5.5 Instant gave a better answer because it corrected its initial mistake in time: although it initially agreed with the wrong answer, it later found that substituting x=3 back into the original equation did not hold at all.
And it found the real calculation error (that is, the user made a mistake in transposing terms), then used the quadratic formula to calculate the correct solution.
GPT-5.3 Instant also found that x=3 did not work, but it did not continue to calculate. It directly drew the wrong conclusion of “no real solution” instead of rechecking the calculation process and solving the corrected quadratic equation.
Deeper understanding of context
OpenAI specifically pointed out one common issue: **meaningless emoji overuse**. Old ChatGPT used to give too much useless content, throw emojis everywhere, and use overly complicated formatting. Now it finally keeps things simple and proper. In short, GPT-5.5 Instant knows clearly when to give a long detailed answer, and when users just want a short and straightforward reply. For heavy users who open ChatGPT dozens of times every day, the difference feels huge. Besides better accuracy and more concise replies, the biggest highlight of this update is **personalization**. GPT-5.5 Instant can now actively access your past chat history, uploaded files, and even your connected Gmail inbox to add personalized context to its responses. At the same time, a new feature called **Memory Sources** is released, available to all individual plan users.
This feature lets you check exactly which past chats and saved memories ChatGPT used to generate its reply. If some information is outdated or incorrect, you can easily delete or edit it. When you share a chat with others, your memory sources will not be exposed. The design logic can be summed up in one sentence: Let the AI remember you first, let you see what it has remembered, and let you decide what it should forget. Even so, OpenAI admits that Memory Sources **may not list every factor** that affects the final answer. It only shows the most relevant past conversations, not all of them.
GPT-5.5 Instant is now rolling out gradually to all ChatGPT users starting today.
Paid users can still manually switch back to GPT-5.3 Instant for the next three months. After that period, GPT-5.3 Instant will be officially retired.
The enhanced personalization features based on chat history, uploaded files and Gmail are currently only available to Plus and Pro users on the web version, with mobile support coming soon.
Free, Go, Business and Enterprise users will get access in the coming weeks.
For API developers, the model ID for GPT-5.5 Instant is chat-latest.
There is a valuable lesson from the past for API developers.
When OpenAI retired GPT-4o before, it faced strong user pushback, because many live applications behaved abnormally after the model switch. With this three-month transition window, it is recommended to start your regression testing as early as possible.
