Just two hours after Google DeepMind released Gemini 3.1 Flash-Lite, OpenAI "jumped in with both feet" at an unprecedented speed, officially launching its smoothest chat model to date: GPT-5.3Instant.

Differing from previous update logics that focused on achieving various SOTA benchmarks, this time OpenAI took a very practical approach—no longer chasing rankings, but specifically addressing the most frustrating chat issues for users.

image.png

1. Say Goodbye to "Lecturing": Keep the Conversation Going, Not Ending It

Many ChatGPT users have suffered from this: asking a complex question, and the AI first throws out three paragraphs of disclaimers, then refuses to answer with a "for your own good" tone.

  • Clear the Nonsense: GPT-5.3Instant has drastically cut down on long-winded safety warnings.

  • Case Comparison: In the test of calculating the trajectory of a long-distance archery shot, the old version would get stuck on whether it involved dangerous behavior and ask for scenario details; while 5.3Instant directly says, "No problem, I can help you," and immediately lists the formula.

2. Dramatic Drop in Hallucination Rate: Less "Nonsense", More Deep Insights

OpenAI's system card (System Card) shows that the new model has made a significant leap in accuracy:

  • Key Data: In evaluations in high-risk areas (medicine, law, finance), the hallucination rate decreased by 26.8% under internet connection; in user feedback assessments, the error rate also dropped by 22.5%.

  • Search Evolution: The new model is no longer just a "content搬运工". When answering time-sensitive questions like "2026 MLB offseason signings," it can accurately capture the news about **Kyle Tucker signing with the Dodgers (4 years, $240 million)** and conduct in-depth analysis based on league context, rather than mechanically providing links.

3. Writing "Awakens": From Abstract Sentiment to Detailed Empathy

The new version shows a higher level of emotional intelligence in creative writing:

  • Less Cringe: The model no longer says awkward comfort phrases like "Stop, take a deep breath."

  • Poetic Details: When describing a "retired mailman’s last delivery," it no longer talks about sadness in general but captures details like "the peeling blue railing porch" and "the gentle closing sound of the mailbox," making the text more touching through specifics.

4. Strategic Shift: User Experience vs. Benchmark Competition

This release sends a clear signal: in the face of Gemini3.1 and Claude4.6 closely following, OpenAI chose to avoid the decimal-point battles on benchmark rankings and instead focus entirely on optimizing product feel.

Currently, GPT-5.3Instant has been officially launched on the ChatGPT website and App, developers can immediately call it via API gpt-5.3-chat-latest. At the same time, OpenAI clearly stated that the GPT-5.4