The most widely used ChatGPT model has just received a major update. The all-new GPT-5.3 Instant is now live, directly targeting users' daily interaction experiences with a focus on optimizing tone倾向, response relevance, and conversational fluidity.
The highlight of this update is that it directly addresses the most strongly felt user experience pain points from before.
Drastically Reduces Refusals and Preachiness
Addressing the model's previous overly conservative safety boundaries, GPT-5.3 Instant has comprehensively optimized its refusal logic. When faced with a query, as long as it can provide useful help, the model will get straight to the point, completely removing past over-defensive opening statements, lengthy restriction explanations, and preachy tones. Communication barriers are greatly reduced, and responses become more direct and effective.
More Natural and Fluid Conversational Style
Previously, the model occasionally exhibited dominant language or over-interpreted user intent without confirmation. This update thoroughly cleans up such dissonance. Mechanical, preachy tones like "stop and take a breath" have been drastically reduced. Not only is the conversational style more natural, but consistency in personality across version updates has also been strengthened. Users can still adjust the model's friendliness and enthusiasm level in settings.
Web Search No Longer Just Lists Links
When processing web information, the new model no longer simply lists and aggregates search results. Instead, it leverages its own knowledge graph to provide deep background analysis for the latest information. This directly solves the occasional problem of link堆积 and scattered information. Now, the model can accurately capture the subtext of questions, prioritize core content at the beginning of answers, all without compromising response speed or tonal performance.
Data: Hallucination Rate Reduced by Over 26% at Highest
In terms of factual accuracy, two core internal evaluations show a significant reduction in hallucination rates:
High-risk domains (covering medical, legal, financial): Hallucination rate dropped by 26.8% in web-search mode and 19.7% using only the model's internal knowledge.
High-frequency hallucination scenarios (based on anonymized user correction logs): Hallucination rate decreased by 22.5% in web-search mode and 9.6% in offline mode.
Advanced Writing Assistance Capabilities
GPT-5.3 Instant's text generation is more resonant, imaginative, and immersive. Whether drafting novels, polishing paragraphs, or exploring inspiration, it can seamlessly switch between practical tasks and expressive writing while maintaining high logical clarity and coherence.
Full Rollout and Future Plans
Starting immediately, GPT-5.3 Instant is available to all ChatGPT users. Developers can call it via API using the model name gpt-5.3-chat-latest. Updates for Thinking and Pro versions will also be released soon.
For the previous model GPT-5.2 Instant, the platform will retain it for three months in the legacy model dropdown for paying users, with a planned full retirement on June 3, 2026.
Reference: