OpenAI's GPT-5.5 Instant becomes ChatGPT default, cuts hallucination by 24%
New default model scores 81.2 on AIME math, up from 65.4
OpenAI has made GPT-5.5 Instant the default model for ChatGPT, replacing GPT-5.3 Instant which launched just two months earlier in March 2026. The new foundation model focuses heavily on accuracy in high-stakes domains: OpenAI says hallucination rates have dropped significantly in medicine, law, and finance, while latency remains low. Benchmark results back this up—GPT-5.5 Instant scored 81.2 on the AIME 2025 mathematics test versus 65.4 for its predecessor, and 76 on MMMU-Pro (multimodal reasoning) versus 69.2. The model also brings improved context management: it can now reference files, previous conversations, and even Gmail via the search tool to generate personalized responses.
Key user-facing features include memory source transparency—users can see where ChatGPT got its information, delete outdated sources, and correct wrong responses. Sharing chats is now possible without exposing memory sources. The model handles repetitive tasks and STEM queries better, and decides when to use web search for more useful answers. Access rolls out first to Plus and Pro subscribers on web, with mobile and Enterprise/Go/Free tiers following in coming weeks. Developers can use GPT-5.5 Instant via API as "chat-latest". This update comes as OpenAI experiments with ads in the US on Free and Go plans. The GPT-5.5 family launched in April 2026 with improved knowledge work and coding abilities.
- GPT-5.5 Instant scores 81.2 on AIME 2025 math vs 65.4 for GPT-5.3 Instant
- Hallucination reduced in medicine, law, and finance; low latency maintained
- New memory source transparency lets users see, edit, and delete sources; share chats without exposure
Why It Matters
A smarter, more reliable default model means professionals can trust ChatGPT for critical research and decision-making.