OpenAI Launches GPT-5.5 Instant with 52% Fewer Hallucinations and Memory Sources Feature

Summary

OpenAI has announced the rollout of GPT-5.5 Instant, a new default model for ChatGPT that demonstrates significant improvements in accuracy and reliability. The model shows 52.5% fewer hallucinations on high-risk topics such as medicine, law, and finance compared to its predecessor GPT-5.3 Instant, with inaccurate claims dropping by 37.3% on previously flagged conversations.

Beyond hallucination reduction, GPT-5.5 Instant delivers substantial benchmark improvements across critical domains. On AIME 2025, a competitive math exam, accuracy jumped from 65.4% to 81.2%, while PhD-level science reasoning (GPQA) climbed from 78.5% to 85.6%. The model also produces tighter, more concise responses without sacrificing substance, reducing unnecessary follow-ups and excessive formatting.

A key feature accompanying the update is "memory sources," which provides transparency into how the model generates personalized responses. Users can now see which stored context—past chats, saved notes, or uploaded files—informed a given response, with granular controls to flag, edit, or delete individual entries. The rollout prioritizes access while managing advanced features: all ChatGPT users gain immediate access to GPT-5.5 Instant, though enhanced personalization is initially limited to Plus and Pro subscribers.

• Model produces shorter, more concise answers with less formatting and fewer unnecessary follow-ups
• Immediate rollout to all ChatGPT users; advanced personalization features phased to paid subscribers over coming weeks