OpenAI unveils GPT-5.3 Instant with fewer refusals

OpenAI has released GPT-5.3 Instant, an update to the most-used model in ChatGPT, aimed at making everyday conversations more fluid and useful.

The update focuses on tone, relevance, and conversational pacing. OpenAI says it should also improve accuracy-especially when using web information-and reduce replies that end in refusals or lengthy caveats.

GPT-5.3 Instant is now available to all ChatGPT users. Developers can access it through the API as gpt-5.3-chat-latest.

Refusals and tone

OpenAI says user feedback showed GPT-5.2 Instant sometimes refused questions it could have answered safely, and that responses on sensitive topics could be overly cautious.

GPT-5.3 Instant reduces what OpenAI describes as unnecessary refusals and cuts back on preambles it characterises as defensive or moralising. OpenAI says the model should answer directly when an answer is appropriate and safe.

OpenAI also says GPT-5.2 Instant could sound "cringe" and make unwarranted assumptions about a user's intent or emotions.

According to OpenAI, GPT-5.3 Instant is more concise, with fewer unnecessary proclamations, and is designed to maintain a more consistent personality across conversations and updates.

ChatGPT users can still adjust response tone, including warmth and enthusiasm, through settings.

Web responses

OpenAI says the model gives stronger answers when drawing on web information and frames the change as a shift away from simply summarising search results.

According to OpenAI, GPT-5.3 Instant is less likely to overemphasise web results, a pattern that could previously lead to long link lists or loosely connected material.

OpenAI also says the model places greater emphasis on the "subtext" of questions. It aims to surface the most important information earlier in a response while maintaining speed and tone.

Accuracy claims

OpenAI says GPT-5.3 Instant reduces hallucinations across a wide range of topics, based on two internal evaluations.

One evaluation covered higher-stakes domains such as medicine, law, and finance. The second measured hallucination rates using de-identified ChatGPT conversations that users flagged as factual errors.

In the higher-stakes evaluation, OpenAI reports that GPT-5.3 Instant reduced hallucination rates by 26.8% when using the web and by 19.7% when relying on internal knowledge, compared with prior models. In the user-feedback evaluation, OpenAI reports a 22.5% reduction with web use and a 9.6% reduction without web access.

Writing use

OpenAI also presents GPT-5.3 Instant as a stronger writing tool, saying it produces more "resonant, imaginative, and immersive prose" and is better suited to drafting fiction, refining passages, and exploring ideas.

OpenAI says it can also move between practical tasks and expressive writing with fewer losses in clarity and coherence.

Known limits

OpenAI highlights remaining limitations in non-English languages, saying responses in some languages-including Japanese and Korean-can still sound stilted or overly literal.

OpenAI says it will continue to monitor feedback on tone and expand customisation options.

Rollout details

OpenAI says updates to ChatGPT products labelled Thinking and Pro will follow soon. GPT-5.2 Instant will remain available for three months for paid users in a Legacy Models section of the model picker.

OpenAI says it conducted safety training and evaluations for GPT-5.3 Instant and published details in a system card.