OpenAI claims ChatGPT’s new default model hallucinates way less
New default model also cuts unnecessary emoji and improves personalization with memory sources.
OpenAI’s GPT-5.5 Instant, the new default model for ChatGPT, boasts significant improvements in factuality, reducing hallucinations by 52.5% on high-stakes prompts covering medicine, law, and finance, according to internal evaluations. In particularly challenging user-flagged conversations, inaccurate claims dropped by 37.3%. The model also delivers tighter, more to-the-point responses with fewer 'gratuitous emojis,' addressing common user complaints about verbosity and style. OpenAI published a system card detailing its evaluation methodology, underscoring a focused effort to tackle AI’s persistent hallucination problem.
Beyond accuracy, GPT-5.5 Instant enhances personalization by pulling context from previous chats and linked services like Gmail, enabling more tailored responses. A new 'memory sources' feature shows exactly what context was used, letting users delete or correct information. This personalization rolls out first to Plus and Pro users on the web, with mobile support coming soon. The model begins rolling out Tuesday to all ChatGPT users, while GPT-5.3 Instant remains available for three months to ease the transition. Free, Go, Business, and Enterprise tiers will gain personalization later, and memory sources is now live on web for consumer plans.
- GPT-5.5 Instant reduces hallucinations by 52.5% on high-stakes prompts (medicine, law, finance) and 37.3% on challenging flagged conversations.
- The model produces tighter, more to-the-point responses and avoids gratuitous emojis.
- New memory sources feature shows context used for personalization, with ability to delete or correct information; enhanced personalization uses previous chats and Gmail.
Why It Matters
For professionals relying on ChatGPT for critical tasks, fewer hallucinations means more trustworthy AI assistance.