2026年9月10日

Is OpenAI Losing Its Edge? Signs the Innovation Icon Is Becoming Ordinary

OpenAI officially launched GPT-5.1 on November 12, introducing two new variants built upon the GPT-5...

OpenAI officially launched GPT-5.1 on November 12, introducing two new variants built upon the GPT-5 family: GPT-5.1 Instant and GPT-5.1 Thinking.

According to the company’s announcement, the Instant model is “more enthusiastic, more intelligent, and better at following user instructions” than its predecessor. The Thinking model, on the other hand, “understands more easily, handles simple tasks faster, and stays more persistent on complex problems.” In most scenarios, ChatGPT will automatically match users to the model that best fits their needs.

OpenAI emphasized that this upgrade is designed from a more user-centric perspective. The goal is not only to make the system smarter, but also more pleasant to interact with. As a result, both its reasoning abilities and conversational style received notable refinements.

Yet the updates hint at a growing criticism: OpenAI is starting to feel a little too… ordinary.

A Step Further from True Intelligence

One major change is the expanded customization of ChatGPT’s tone. Developers now have more “intuitive and effective” options to shape how the chatbot speaks, allowing users to choose from an array of predefined styles.

While tone presets were introduced earlier this year, the latest update adds new options such as “Professional,” “Candid,” and “Quirky,” while keeping refined versions of the original “Default,” “Friendly” (formerly “Listener”), and “Efficient” (formerly “Robot”).

Beyond presets, OpenAI now lets users fine-tune ChatGPT’s personality directly in the settings—adjusting response length, warmth, readability, and even emoji frequency for more precise control over the chatbot’s behavior.

After months of criticism that the model sounded too “flattering,” OpenAI has now fully embraced that label—handing users the ability to choose exactly how flattering (or not) their chatbot should be.

At first glance, it sounds thoughtful. But compared with previous upgrades that felt groundbreaking, this update delivers less of that “wow” factor. Users ultimately expect AI to deliver performance—not just personality. Productivity relies on utility, not merely emotional tone. And when efficiency gains are marginal, the technology drifts further from true “intelligence,” inching closer to entertainment or novelty value.

Interestingly, ChatGPT can now also adjust tone on the fly. If users feel overwhelmed by settings, they can simply state their preference mid-conversation, and the system will adapt without needing to visit the preferences panel.

Below is a closer look at what GPT-5.1 brings to the table.

“Understands Human Language Better”

OpenAI describes GPT-5.1 Instant as more natural, friendly, and conversational by default. Compared with the GPT-5 model released in August, the official examples highlight improvements in phrasing and clarity, though the enhancements focus more on polish than substance.

A more meaningful upgrade appears in instruction obedience. For instance, when users request replies “always in six words,” the new model responds more consistently and concisely than the previous version.

While not perfect, it does show a noticeable improvement in understanding directives—especially compared with older models that often misunderstood or ignored them.

OpenAI also stated that GPT-5.1 Instant is the first to use adaptive reasoning, allowing it to autonomously decide when to “think” before answering challenging questions. This enables faster responses while still delivering accurate, detailed answers. It reportedly performs well in evaluations like AIME 2025 and Codeforces.

“Strong When Needed, Fast When Possible”

GPT-5.1 Thinking is even more flexible. It dynamically adjusts its reasoning depth depending on task complexity—offering faster responses for simple questions and more elaborate answers for difficult ones.

Benchmark results show that with “standard” thinking time, the Thinking model outperforms GPT-5 on fast tasks, yet also takes longer on slow tasks. In other words, it becomes stronger when the question demands it, and lighter when it doesn’t—truly ‘strong when strong is needed, weak when weak is enough.’

OpenAI noted that the Thinking model uses fewer technical terms and ambiguous phrases, making explanations clearer and more digestible—particularly in professional settings or when breaking down complex concepts.

In an example explaining “BABIP and wRC+,” the model’s upgraded version not only maintains the same informational depth but also adds formula explanations and guiding context—resembling a polished “instructor’s edition” of a textbook.

These two new models will roll out to ChatGPT users starting this week. To prevent another backlash like the last “cliff-drop” experience, OpenAI will retain the previous GPT-5 model under a “Legacy Models” menu for three months before removing it.

Overall, this tone-focused update feels somewhat “concept stock”—aimed more at market sentiment than technological advancement. OpenAI seems to be responding to emotional preferences, quietly drifting from its core trajectory of pushing technical frontiers. As AI becomes more commercial, it is learning to “speak nicely,” shifting toward a form of surface-level intelligence.

But true intelligence shouldn’t just be about being “better at chatting.” It should understand the world more deeply, reason more robustly, and solve real, complex problems.

接著讀