GPT-5.5 Instant replaces the existing GPT-5.3 Instant and is also available via the API as "chat-latest." According to OpenAI, the update brings improvements in factual accuracy, response precision, and personalization.
Fewer Hallucinations on Sensitive Topics
In internal evaluations, GPT-5.5 Instant produced 52.5% fewer hallucinated claims than its predecessor when handling high-risk prompts in areas such as medicine, law, and finance. For particularly difficult conversations that users had previously flagged for factual errors, inaccurate claims dropped by 37.3%.
OpenAI illustrates the difference with an algebra example: a user uploaded a photo of a handwritten equation that contained a calculation error. GPT-5.3 Instant recognized that the proposed solution x=3 didn't work out, but incorrectly concluded there was no real solution. GPT-5.5 Instant went further: while it also initially agreed with the user's calculation, it then continued reasoning, identified an error in the user's rearrangement, and solved the corrected quadratic equation using the quadratic formula.
The improvements are also reflected in benchmark results. On AIME 2025 (a math competition), accuracy rose from 65.4% to 81.2%. On GPQA, a PhD-level science test, the model improved from 78.5% to 85.6%. Visual reasoning (CharXiv) climbed from 75.0% to 81.6%, multimodal expert knowledge (MMMU-Pro) from 69.2% to 76.0%, and document parsing error rates (OmniDocBench) fell from 14.6% to 12.5%.
| Benchmark | Description | Metric | GPT-5.3 Instant | GPT-5.5 Instant |
|---|---|---|---|---|
| CharXiv-reasoning | Scientific Chart Reasoning | Accuracy | 75.0% | 81.6% |
| MMMU-Pro | Expert Multimodal Reasoning | Accuracy | 69.2% | 76.0% |
| OmniDocBench | Document Parsing | Avg. error rate (lower = better) | 14.6% | 12.5% |
| GPQA | PhD-Level Science | Accuracy | 78.5% | 85.6% |
| AIME 2025 | Competition Math | Accuracy | 65.4% | 81.2% |
More Concise and More Personal Responses
Alongside accuracy, OpenAI says it has focused on conciseness. Responses are intended to be tighter without sacrificing substance. The model asks fewer unnecessary follow-up questions, skips redundant emojis, and avoids over-formatting. The tone is informal and practical, without over-explaining.
GPT-5.5 Instant is also designed to draw more effectively on context from past chats, uploaded files, and connected Gmail accounts — provided users have enabled these features. The model makes smarter decisions about when a response benefits from additional personalization and searches through previous conversations more quickly.
Alongside this, OpenAI is introducing Memory Sources across all ChatGPT models. When a response has been personalized, users can now see for the first time exactly what context was used — such as saved memories or previous chats. Individual entries can be marked as relevant or irrelevant, edited, or deleted.
OpenAI notes, however, that Memory Sources may not display every factor that influenced a response — for example, only a portion of searched chats will be listed as sources. The company plans to make this view more comprehensive over time. Memory Sources are not shared when a chat is shared with others. Temporary chats neither access nor update memory.
Staged Rollout Across Plans
GPT-5.5 Instant is being rolled out to all ChatGPT users starting immediately. For paying users, GPT-5.3 Instant will remain accessible via model settings for another three months before being retired.
Extended personalization through past chats, files, and Gmail is initially available to Plus and Pro users on the web, with mobile support coming soon. Expansion to Free, Go, Business, and Enterprise plans is planned over the coming weeks. Memory Sources will roll out for all consumer plans on the web, with mobile availability to follow. Certain personalization sources may vary by region.
ES 
EN