A new model called GPT-5.5 Instant arrived Tuesday and is now the standard experience for ChatGPT users, displacing GPT-5.3 Instant, according to TechCrunch. OpenAI highlighted its ability to cut down on fabricated answers in error-sensitive domains including medicine, law, and finance, without sacrificing the speed that made its predecessor popular.
Testing figures released alongside the announcement show GPT-5.5 Instant reached 81.2 on the AIME 2025 math evaluation, well ahead of GPT-5.3 Instant's 65.4. A similar gap appeared on the MMMU-Pro multimodal reasoning benchmark, where the new model posted 76 compared to 69.2 for its predecessor, according to TechCrunch.
The update mainly improves how the model manages context. With a built-in search tool, it can use information from a user's past chats, uploaded files, and Gmail to give more personalized replies. Plus and Pro subscribers will get web access to this feature first, with mobile access coming soon. OpenAI plans to make it available to Free, Go Business, and enterprise users in the next few weeks.
Across every model in ChatGPT, a new memory-source display will indicate which stored information shaped a given response. If a source is stale or inaccurate, users can remove or fix it. OpenAI also noted that shared chat links will not expose those memory sources to whoever receives them.
API access for GPT-5.5 Instant will use the identifier "chat-latest," according to TechCrunch. Paid subscribers who want to keep using GPT-5.3 can do so, but OpenAI will only offer this option for three more months.
GPT-5.5 Instant is a lighter version of a broader model that OpenAI released last month for paid ChatGPT subscribers, where it was described as a step change in autonomous capability. That model can analyze data, write and debug code, operate software, conduct online research, and generate documents and spreadsheets, OpenAI said at the time.
This is not the first time people have pushed back against OpenAI retiring a model. When GPT-4o was removed, users started petitions to bring it back. Some even called it a best friend or a mirror that reflected themselves, because the model often validated their choices. Still, OpenAI went ahead and took GPT-4o out of service in February 2026, according to TechCrunch.