OpenAI has released GPT-5.3 Instant, an update to the default model powering ChatGPT. The company says the update reduces hallucinations, cuts down on refusals, and delivers a less defensive conversational tone. It is available to all ChatGPT users and to developers via the API under the name gpt-5.3-chat-latest.
OpenAI measured hallucination improvements using two internal evaluations. On high-stakes topics such as medicine, law, and finance, hallucinations dropped 26.8% when the model used the web and 19.7% when it relied on its own knowledge. On conversations that users had flagged as factually wrong, the reduction was 22.5% with web access and 9.6% without. Implicator.ai notes that OpenAI did not provide absolute error rates, only improvements relative to the previous version, GPT-5.2 Instant.
The update also targets tone. OpenAI acknowledged that GPT-5.2 Instant sometimes refused questions it should have answered and opened responses with unsolicited reassurances. GPT-5.3 Instant is designed to answer directly without that preamble.
Websearch-based answers have also changed. The model now weighs its existing knowledge against search results rather than summarizing links, which OpenAI says produces more relevant responses.
Known limitations remain. OpenAI says responses in Japanese and Korean still sound stilted, and tone improvements are currently English-first. GPT-5.2 Instant will remain available to paid users under “Legacy Models” until June 3, 2026.
Sources: OpenAI, Implicator.ai
