OpenAI said on the 4th that it improved the features of its flagship ChatGPT model, "GPT-5.3 Instant," reducing hallucinations by up to 27%.
OpenAI said, "With this update, the answer accuracy of 'GPT-5.3 Instant' has improved, and the conversation experience is smoother than the previous version." The company said it focused on boosting the accuracy of fact-based responses across a range of topics, which in turn reduced hallucinations—when AI makes up plausible but untrue answers.
In particular, in internal evaluations targeting high-risk areas such as law and finance, hallucination rates fell 26.8% when using web search and 19.7% when relying only on internal knowledge. In evaluations based on de-identified user feedback, hallucinations dropped 22.5% with web use and 9.6% without the web.
The conversation style and flow were also refined. It significantly reduced the "cringe" noted as a problem in the previous version and unnecessary refusals to answer, and it improved responses that were overly emotional or that hastily inferred user intent.
The way it uses web information has also been improved. "GPT-5.3 Instant" more effectively combines information found online with its own knowledge and reasoning, going beyond merely summarizing search results to explain the context of recent news or issues based on existing knowledge. It also better identifies the hidden intent of questions and presents key information at the start of the answer to deliver more relevant, immediately actionable responses.
Writing support features were also strengthened. It is designed to support richer, more immersive writing across various creative tasks, including drafting novels, editing sentences, and exploring ideas.
"GPT-5.3 Instant" is available to all ChatGPT users. Developers can access it as the "gpt-5.3-chat-latest" model via the application programming interface (API). OpenAI plans to roll out updates to the "GPT-5.3 Thinking" model and the Pro model in phases.
Meanwhile, the existing "GPT-5.2 Instant" will be offered to paid users for three months in the "legacy models" section of the model selector, and the service will end on Jun. 3.