OpenAI Releases GPT-5.3 Instant With Fewer Refusals & Lower Hallucination Rates

OpenAI has released GPT-5.3 Instant, an update to the most widely used model in ChatGPT, to improve everyday conversations, reducing unnecessary refusals and lowering hallucination rates.

In a statement announcing the update, the company said the model delivers “more accurate answers, richer and better-contextualised results when searching the web, and reduces unnecessary dead ends, caveats, and overly declarative phrasing that can interrupt the flow of conversation.”

The update focuses on improving tone, relevance, and conversational flow—areas OpenAI said are frequently raised in user feedback. “GPT-5.3 Instant directly reflects user feedback in these areas,” the company said.

GPT-5.3 Instant is available immediately to all ChatGPT users and to developers through the API under the name ‘gpt-5.3-chat-latest’.

OpenAI said GPT-5.2 Instant will remain available for paid users in the model picker under the Legacy Models section for three months before being retired on June 3.

OpenAI said the new model reduces unnecessary refusals and limits defensive or moralising responses when addressing sensitive topics. The goal is to provide answers directly when a question can be addressed safely.

“GPT-5.3 Instant significantly reduces unnecessary refusals, while toning down overly defensive or moralising preambles before answering the question,” the company said. “When a useful answer is appropriate, the model should now provide one directly.”

The company also changed the model’s conversational style. According to OpenAI, the previous version could sound “overbearing” or make assumptions about a user’s intent.

OpenAI also reported improvements in factual accuracy. On internal tests covering areas such as medicine, law, and finance, the model reduced hallucination rates by 26.8% when using web search and 19.7% without web access, compared with earlier models.

On another internal evaluation based on conversations flagged by users for factual errors, hallucinations declined 22.5% with web access and 9.6% without it, the company said.

OpenAI also said GPT-5.3 Instant improves writing capabilities. The company said the model performs better in tasks such as drafting fiction, editing text, and generating longer written responses.

“GPT-5.3 Instant is also a stronger writing partner,” OpenAI said, adding that the model can move between practical tasks and creative writing “without losing clarity or coherence.”

The company acknowledged several limitations. It said responses in some languages, including Japanese and Korean, may still sound “stilted or overly literal,” and that improvements in tone and naturalness across languages remain ongoing.