OpenAI released GPT-5.3 Instant on 3 March 2026, an update to the default ChatGPT model aimed at producing more natural conversations and more accurate answers. The company framed the release around a complaint that does not show up in benchmark scores but strongly shapes how people feel about a model: that earlier versions could sound stilted, preachy or, in OpenAI’s own word, “cringe.”
What changed
In an accompanying blog post, OpenAI said the update focused on “tone, relevance, and conversation flow,” following feedback that the previous model, GPT-5.2 Instant, sometimes declined questions it could safely answer and replied in an overly cautious or sermonising way. The company acknowledged that those traits could make the model feel overzealous, and said the new version trims unnecessary preambles, lengthy warnings and abrupt dead ends so that answers read more smoothly.
Accuracy gains
OpenAI also reported measurable reductions in hallucination — confident statements that are factually wrong. On higher-stakes evaluations spanning areas such as medicine, law and finance, the company said GPT-5.3 Instant cut hallucination rates by about 26.8 percent when drawing on web search and 19.7 percent when relying only on internal knowledge. On evaluations built from user-flagged feedback, it reported drops of roughly 22.5 percent with web access and 9.6 percent without. These are OpenAI’s own internal measurements, so independent testing remains the firmer guide, and hallucination in large language models is a structural problem rather than a solved one, as this explainer describes.
Where it did not improve
The update was not uniformly better. OpenAI noted a small regression on parts of its safety benchmarking — specifically around disallowed sexual content relative to versions 5.2 and 5.1, and self-harm content relative to 5.2 — and said it was monitoring those areas, cautioning that post-launch benchmark values can shift. The company also flagged remaining work on tone in other languages, observing that the model could be especially verbose in Japanese and Korean, alongside plans to expand customisation options.
What it means in practice
For everyday users and for businesses building on the model, the trade-off is familiar: a friendlier, less hedging assistant is more pleasant to use, but reduced caution has to be balanced against safety. Teams that rely on these models for customer-facing work — for example through no-code AI agent builders — are well advised to test updated models against their own tasks rather than assume an upgrade is strictly better on every dimension.
Already superseded
The pace of change is its own caveat. GPT-5.3 Instant was itself replaced as the default ChatGPT model in May 2026 by GPT-5.5 Instant, which OpenAI said reduced hallucinated claims substantially further on high-stakes prompts. The broader pattern is steady, incremental tuning of tone and reliability rather than a single decisive leap — a reminder that model version numbers move quickly and that the “current” default is a moving target.
