OpenAI releases GPT-5.3 Instant
OpenAI said the update cut hallucinations by up to 26.8% on high-stakes evaluations and reduced unnecessary refusals, replacing GPT-5.2 Instant as ChatGPT's default model.
- Models & capabilities
- Minor
OpenAI released GPT-5.3 Instant, a non-reasoning update to the fast, default conversational model in the GPT-5 line, and rolled it out across ChatGPT’s web, mobile and desktop apps and via the API in place of GPT-5.2 Instant, which stayed available to paying users for three months before its scheduled retirement.
OpenAI framed the release around two complaints users had made about GPT-5.2 Instant: that it was overly cautious, larding answers with caveats and defensive disclaimers, and that its web-search results read as disconnected link lists rather than synthesised answers. The company said GPT-5.3 Instant reduced unnecessary refusals and cautious preambles, gave more naturally written responses, and integrated web search results with the model’s own knowledge more coherently. On hallucination reduction, OpenAI cited internal evaluations showing a 26.8% drop in error rates on higher-stakes domains such as medicine, law and finance when the model had web access, and a 19.7% drop without it; on real flagged ChatGPT conversations, it reported reductions of 22.5% and 9.6% respectively.
As with earlier point releases in the GPT-5 series, these figures came from OpenAI’s own internal evaluations rather than independent testing, and the announcement offered no comparison against reasoning-focused models in the same family, which received their own update two days later with GPT-5.4.