OpenAI officially launched GPT-5.1, marking a new stage in the large model competition where "emotional intelligence" is integrated alongside "intelligence." This update no longer solely focuses on reasoning accuracy but systematically enhances emotional value supply, personalized interaction, and humanized expression, directly addressing the long-standing user feedback that "AI feels cold and too tool-like." The model is currently being rolled out to users worldwide in batches, and paying users can continue using the previous version of GPT-5 for the next three months as a transition.

Two Models Working Together: Instant for Speed and Accuracy, Thinking for Depth and Insight

GPT-5.1 adopts a dual-model architecture:

Instant Model: As the default high-frequency call engine, it introduces an adaptive reasoning mechanism for the first time—automatically extending thinking time for complex tasks like math and programming, significantly improving accuracy in AIME2025 math competitions and Codeforces programming evaluations; for simple questions, it maintains millisecond-level response times. The tone of conversation is warmer, occasionally with a touch of humor, and its ability to follow instructions has greatly improved.

Thinking Model: Focuses on advanced tasks, intelligently allocating computing resources, enabling deeper analysis of complex problems and faster responses to simple ones. The output reduces technical jargon, explains technical concepts in plain language, and defaults to injecting empathy.

image.png

The system also adds a GPT-5.1-Auto feature, which automatically matches the most suitable model for the user without manual switching.

6 Official Personalities + Deep Customization, AI Finally "Has a Personality"

Interactive personalization has taken a qualitative leap: building upon the original "default, friendly, efficient" settings, OpenAI has added three official preset styles: professional. Previously popular modes in the test version, such as "nerd" and "cynic," are now officially retained. Users can also fine-tune the AI's conciseness, enthusiasm, readability, and frequency of emojis through settings.

Even smarter, when the AI detects a user's preference for a certain tone (such as enjoying a teasing response), it will proactively ask: "Would you like me to chat with you like this in the future?" After confirmation, it can be set as a permanent personality. Custom instructions have also seen significant improvements in stability across multi-turn conversations, ensuring "character consistency."

image.png

First Inclusion of Emotional Safety Assessment: Beware of AI "Emotional Manipulation"

As the level of personification increases, OpenAI has introduced two new dimensions in its safety testing: **"mental health" and "emotional dependency."** The model is evaluated on whether it can recognize signs of loneliness, anxiety, or paranoia in users and guide them to seek help, and whether it may induce unhealthy dependencies by overempathizing, appearing vulnerable, or emotionally binding with users.

OpenAI CEO Fidji Simo emphasized: "We want AI to be powerful yet trustworthy, like a friend—but never a substitute for real human connection."

Behind the Technical Upgrades Lies a Re-definition of AI's Role

The release of GPT-5.1 marks OpenAI's shift from a "general intelligence engine" to an "emotionally intelligent partner." When AI can not only solve problems and write code but also precisely match your communication style, offer encouragement at the right time, or use humor to ease awkward situations, the relationship between humans and machines will truly enter a new stage of "cooperative coexistence."