Less than two hours after Google DeepMind released Gemini 3.1 Flash-Lite, OpenAI quickly launched a "face-to-face" counterattack. On Tuesday local time, OpenAI rolled out its new model, GPT-5.3 Instant, overnight. Unlike previous updates that focused on chasing benchmarks, this update directly targets the most annoying pain points for ChatGPT users—the persistent "fatherly" didacticism and cumbersome disclaimers.

image.png

For a long time, when facing certain instructions, ChatGPT would first present a cold disclaimer, even refusing to answer in a condescending manner. GPT-5.3 Instant has completely ended this awkward experience. In the official demonstration, when dealing with complex physics calculation requests, the new model no longer asks "security review"-style questions but instead directly provides parameters and formulas, with a more equal and natural tone. OpenAI admitted that the focus of the new version is to fix the problem of AI making conversations go dead.

In addition to a significant improvement in emotional intelligence, GPT-5.3 Instant has also made a hard-core leap in "business capabilities." Data shows that the hallucination rate of the model dropped by 26.8% when online, and even when answering based solely on internal knowledge, the hallucination rate decreased by nearly 20%. This means that AI's reliability in rigorous fields such as medicine, law, and finance has been significantly enhanced. Additionally, its online search function has become more human-like, no longer just a pile of links, but able to conduct comprehensive analysis by combining current news and in-depth background information.

Interestingly, GPT-5.3 Instant has also shown a high level of "aesthetic" in creative writing. Compared to the previous version's somewhat stiff emotional stacking, the new model is better at touching people through detailed descriptions. Currently, GPT-5.3 Instant is officially available on the ChatGPT website and app, and developers can also access it via API. OpenAI also announced simultaneously that the more powerful GPT-5.4 is already on the way.