OpenAI Releases “GPT-5.3 Instant”: A “Less Cringe” Model Free from Unnecessary Lecturing and Refusals
On March 3, 2026, OpenAI released “GPT-5.3 Instant,” updating the default model for ChatGPT and significantly improving the quality of everyday conversations.
Rather than simply adding new features, this update focuses on addressing user complaints about “overly defensive tones” and “unnecessary refusals,” delivering more natural and direct interactions.
“Less Cringe” Improvements to Eliminate Stress in Everyday Conversations
The primary theme of GPT-5.3 Instant is reducing frustrating or “cringe-worthy” interactions for users.
- Elimination of Preachy Tones: The previous model (GPT-5.2 Instant) had a tendency to misinterpret user intentions and output overly moralistic, pushy phrases like “Stop. Take a breath.” In the new model, these dramatic expressions have been cut, resulting in a much more natural conversational flow.
- Reduction of Unnecessary Refusals: The issue of the model applying overly strict safety standards to reject safe questions—or delivering lengthy caveats before answering—has been resolved. The new model skips the preambles and provides direct answers.
Reduced Hallucinations and Enhanced Web Search
In addition to the conversational tone, the accuracy of information and web search integration capabilities have also been improved.
- Lower Hallucination Rates: According to OpenAI’s internal evaluations, the hallucination rate has significantly decreased compared to previous models. A 26.8% reduction in false information (22.5% in user evaluations) was confirmed when using web search, and a 19.7% reduction (9.6% in user evaluations) was noted even when relying solely on the model’s internal knowledge.
- Proper Contextualization of Information Sources: When utilizing web search, the model no longer merely summarizes and lists found information. Instead, it balances the search results with its own knowledge, prioritizing the most important information in alignment with the user’s intent.
Enterprise Rollout and Future Schedule
GPT-5.3 Instant is already available to all ChatGPT users. The older model, GPT-5.2 Instant, is scheduled to be retired on June 3, 2026, following a three-month transition period.
For enterprise users, the model will soon be rolled out as “GPT-5.3 Chat” viaMicrosoft Foundry. Pricing is set at
1.75per1millioninputtokens(1.75per1millioninputtokens(0.175 for cached inputs) and $14.00 per 1 million output tokens, supporting stable operations for large-scale customer support and internal IT workflows.
Looking ahead, OpenAI plans to release “GPT-5.3 Pro” and “GPT-5.3 Thinking” models, which are specialized for more complex reasoning tasks, further expanding the GPT-5.3 family.
