OpenAI has released GPT-5.3 Instant with major improvements focused on real-world usage: reduced unnecessary refusals, better accuracy, and more natural conversations
Kawin Suangkaew

OpenAI has released GPT-5.3 Instant with major improvements focused on real-world usage: reduced unnecessary refusals, better accuracy, and more natural conversations
OpenAI has released an update to their most-used ChatGPT model, focusing on everyday usability improvements:
One of the most significant changes is the reduction in factual errors (hallucinations). OpenAI tested the model in two key areas:
The results are impressive: GPT-5.3 Instant reduces hallucination rates by 26.8% with web search and 19.7% with internal knowledge. In real-world usage from user feedback, the numbers are 22.5% and 9.6% respectively.
This update focuses on the "feeling" of conversation rather than just benchmark numbers, making ChatGPT feel more natural and useful in real life
Many users felt that previous GPT-5 versions were overly cautious or used awkward phrases like "Stop. Take a breath before I answer," which made conversations feel unnatural.
GPT-5.3 Instant fixes this by:
While GPT-5.3 Instant has many improvements, OpenAI acknowledges there's more work to do:
For software developers, this update is significant:
GPT-5.2 Instant will remain available for paid users for 3 more months before being retired on June 3, 2026
GPT-5.3 Instant shows the direction OpenAI wants to take ChatGPT: becoming a more "natural" assistant that's not just great at benchmarks but actually useful in everyday life.
The reduction in hallucinations and making conversations feel more natural is a significant step toward building trust in AI for both general users and developers.