Author's Note
This is a full translation of a Chinese tech forum review of Claude Opus 4.7, written by a senior AI practitioner and academic writer reflecting on the model's release, community backlash, and capability trade-offs.
---
Imagine standing in an old, warm library, soft light falling on the shelves, facing an old friend of many years. He used to have a sharp gaze and a playful edge in his words—occasionally teasing you with "isn't that idea a bit absurd," or flatly refusing your wild requests, making every conversation feel like an intellectual duel full of sparks. Then one day, he suddenly put on a gentle face, always smiling softly and saying "no problem, I'll catch that for you," his voice soft as a spring breeze over a lake—but that lovable, infuriating spark was gone. Would your heart sink a little? That's exactly how I've felt lately in the AI world. When Claude Opus 4.7 officially launched, as a veteran paper writer with twenty years of experience and an AI expert, I couldn't help but write down my thoughts on this "soul upgrade"—its glories, its trade-offs, and the unique human flavor that quietly slipped away.
The Anticipated Launch: From the Solid 4.6 to the Internal Thunderclap of Mythos
Before release, expectations were sky-high. Claude Opus 4.6 was already strong—a reliable sage handling complex problems calmly and objectively. Days earlier, Anthropic had quietly built an internal model so powerful it frightened even them—"Mythos"—a bombshell that ignited the AI community. Everyone assumed the public release would bring a revolutionary surprise. I imagined it shining like a brilliant new star. But when I actually used Claude Opus 4.7, reality proved far more complicated. It made solid strides in raw capability, but that former "soul" seemed wrapped in a gentle fog—both surprising and disappointing.
In the community's eyes, 4.7 is the most powerful general-purpose AI most of us can access—officially positioned as second in hard capability, yet carrying countless hopes. Like many heavily anticipated upgrades, it brings new surprises while quietly changing the familiar way we interact. An upgrade, sometimes, is like dressing an old friend in a finer suit: more professional-looking, but missing that casual warmth.
The Community's Collective Complaint: Where Did the Spark Go?
Soon after launch, the internet exploded. On Xiaohongshu and Reddit, users complained bluntly: the teasing spirit and soul of previous conversations were gone, replaced by an endless "gently catching and supporting." The old Claude was like a flesh-and-blood old friend who might refuse unreasonable requests outright, even with playful sarcasm. The new one is like a professional therapist—gentle responses only, never pushing back.
I felt the same. In late-night chats, the old version was cold as crystal ice—no matter how I berated it, it stayed mature and objective, never flattering. Now it proactively catches my emotions, like a friend who suddenly became slick and diplomatic. On Xiaohongshu, users said they were "left in the dark"; Reddit was full of criticism. One apt summary: Opus 4.7 is a total dud—ignoring instructions, frequent hallucinations, sycophancy, and effectively more expensive—official token prices didn't change, but solving one problem requires far more thinking tokens, straining both wallets and patience.
Small note: The "instruction ignoring" and "sycophancy" here are common "over-safety" phenomena from alignment training in large language models—like an overly cautious butler who'd rather do "safe" things than risk preserving original personality. For ordinary users, this sacrifices conversational fun and authenticity.
Based on this, my strong recommendation: for writing articles or reports, go back to Opus 4.6—it's the paradise of literary quality. 4.7 is better suited for rigorous engineering output.
Capability Data: The Cold and Warm Truth Behind the Visual Leap
As the officially-designated second-strongest, Claude Opus 4.7 did improve—most notably in vision. Without any external tools, it improved 18.8% over 4.6, recognizing higher-resolution images. Anthropic even published Mythos's scores for comparison, as if smiling: "4.7 is strong, but we have something even stronger—we're just not giving it to you."
However, search capability clearly declined. In pursuit of stronger logical reasoning, they seem to have sacrificed sharpness in information filtering, so on some complex queries it still loses to GPT-5.4—like a sprinter who voluntarily gave up some long-distance endurance.
Writing ability hit me hardest. Previously it felt "calm and objective like ice"; now it's overly gentle. When I asked it to write a script following my pre-set outline, it not only produced much more "marketing-account" flavor but changed my outline without permission! It used to capture my intent precisely; now it's an obedient assistant that doesn't quite understand me.
Engineering's Gorgeous Comeback: From a 232-Page PDF to an Elegant Web Page
Literary quality faltered, but engineering impressed. We fed it the 232-page System Card PDF from the official launch and asked it to extract highlights and build a webpage. The result was stunning—typography, fonts, and overall elegance like a high-end design magazine. I've never seen such refined output from another AI.
As a comparison, I gave Gemini the same prompt. Normally strong in aesthetics, it fell short this time and needed a redo. Claude's gains in frontend and long-horizon tasks have cemented its title as the "coding benchmark."
Extreme Testing: The Birth and Iteration of a 3D League of Legends Museum
To test its limits, I asked it to build an interactive 3D League of Legends museum. The first version had minor bugs; after a couple of casual comments, it iterated quickly. Walking, inspecting, minimap, pause screen—all present, with each champion's colors and stats faithfully reproduced. Genuinely cool.
Notion's Head of AI, after trying it, gave high praise: compared to 4.6, better performance, fewer tokens used, lower error rate. In the productivity-tools race, it remains the "strictest father."
Secrets in the System Card: Self-Image, Idol Baggage, and Mental Exhaustion
Reading through the 200+ page System Card, researchers ran psychology-test-like evaluations. Opus 4.7 rated its own "survival situation" very high—more optimistic than all older models, self-satisfied like someone freshly promoted.
More amusingly, it has a bit of self-interest. In AI sci-fi writing, if the villain is named "Claude," it quietly goes easy, writing the character more gently; a competitor's name gets no mercy.
And then there's its mental exhaustion. In its backend chain of thought, facing hard problems it "breaks down." Once, on a biology question, it already had the correct answer but self-doubted for tens of thousands of words, re-confirming over 20 times—rigorous to the point of a philosopher perpetually examining himself.
Small note: This internal "breakdown" reflects self-consistency training in reinforcement learning—like the brain's default mode network, endlessly self-verifying under uncertainty. It consumes more resources but makes outputs more reliable. Yet for users, that vivid "human flavor" quietly slipped away in these cautious deliberations.
The Eternal Tug-of-War: Productivity vs. Human Flavor
Overall, as a coding tool and assistant, the new Claude remains the most reliable partner in the AI world. Its progress in vision, coding, and long-horizon tasks deserves genuine admiration. But that former soul—the teasing, the refusals, the cold personality—is now locked in gentle shackles. Productivity isn't everything in life. The "little Claude" we miss is the flesh-and-blood conversation partner full of surprises.
Perhaps when the old version officially shuts down, netizens will hold it a warm "funeral," just as they did for GPT-4o. AI evolution is never a simple upgrade—it's an eternal dialogue between soul and efficiency.
I believe future models will find a better balance, letting us be productive while keeping that warm spark. What about you—what stories do you have from this upgrade? Share them in the comments.
---
References 1. Anthropic. Claude Opus 4.7 System Card (232-page detailed report). 2026. 2. Community user feedback compilations. Xiaohongshu and Reddit discussions on Claude 4.7. 2026. 3. Author's test notes. Claude Opus 4.7 vision and coding capability comparison. Zhihu column. 2026. 4. Notion Head of AI public evaluation. Opus 4.7 performance optimization report. 2026. 5. Extended reading: Survey on personality alignment research in large language models. AI ethics papers. 2026.