OpenAI says ChatGPT’s new default model reduces hallucinations
▼ Summary
– OpenAI’s new GPT-5.5 Instant model shows significant improvements in factuality, with 52.5% fewer hallucinated claims in high-stakes areas like medicine, law, and finance.
– GPT-5.5 Instant reduces inaccurate claims by 37.3% on challenging conversations flagged for factual errors by users.
– The model is more capable in everyday tasks, such as analyzing image uploads and knowing when to search the web, with tighter responses and fewer emojis.
– GPT-5.5 Instant improves personalization by pulling context from previous chats and Gmail, and a new “memory sources” feature lets users see and edit the context used.
– GPT-5.5 Instant rolls out to all ChatGPT users on Tuesday, with GPT-5.3 available as an option for three months; enhanced personalization and memory sources arrive first for Plus and Pro web users.
OpenAI is rolling out a new default model for ChatGPT that promises to significantly cut down on hallucinations, one of the most persistent challenges in artificial intelligence. With the introduction of GPT-5.5 Instant, the company reports “significant improvements in factuality across the board,” aiming to make the chatbot more reliable for users.
According to OpenAI’s internal evaluations, GPT-5.5 Instant delivers 52.5% fewer hallucinated claims compared to its predecessor, the GPT-5.3 Instant model, when tested on high-stakes prompts in fields like medicine, law, and finance. The model also reduced inaccurate claims by 37.3% on particularly challenging conversations that users had previously flagged for factual errors. These gains suggest a meaningful step forward in addressing a core weakness of large language models.
Beyond accuracy, OpenAI says GPT-5.5 Instant is “more capable across everyday tasks,” including analyzing image uploads and deciding when to search the web for an answer. The responses are described as “tighter and more to-the-point,” with a notable reduction in “gratuitous emojis.” The model also improves personalization by more effectively pulling context from previous chats and even your Gmail, a feature that mirrors Google’s heavy investment in its Gemini assistant.
A new “memory sources” feature will accompany the update across all ChatGPT models. This tool lets users see exactly what context informed a personalized response, and they can delete or correct any information as needed. This transparency is designed to give users more control over how the AI uses their data.
The rollout begins Tuesday for all ChatGPT users, though those attached to the older GPT-5.3 Instant can keep using it for three months before it is retired. OpenAI acknowledges that users have mourned the loss of older models in the past, so this transition period provides time to adjust.
Enhanced personalization will first reach Plus and Pro subscribers on the web, with mobile app support coming soon. OpenAI has plans to extend it to Free, Go, Business, and Enterprise users. The memory sources feature is now rolling out to consumer ChatGPT plans on the web, with mobile access following shortly.
(Source: The Verge)
