[KongchangAI]
· 2 min read· 1,274 words

Gemini 4 Carbon Debuts at No. 1 on Arena; Chinese Models Hold Their Ground on Video Leaderboard

Gemini 4 Carbon Debuts at No. 1 on Arena; Chinese Models Hold Their Ground on Video Leaderboard

Gemini 4 Carbon debuts at No. 1 on Arena; Chinese models hold the video leaderboard; Reddit traffic drops 13% due to AI summaries.

The week 46 Arena blind-test leaderboard shows Google's Gemini 4 Carbon debuting at No. 1 with 1,526 points, followed closely by its sibling Gemini 4 at 1,522, with Google sweeping the top two and forming dual high-position competitiveness. Under the blind-test mechanism, scores reflect real user preferences, but debut scores still need multiple weeks of verification. Chinese models lost the top spot on the overall leaderboard but held firmly in the top five on the video leaderboard, reflecting a differentiation strategy. Meanwhile, SimilarWeb data shows Reddit's global traffic fell 13% over the year, as Google's AI Overviews provide answers directly and reduce search referrals, and Reddit plans to require visitors to log in to cope with changes in the traffic structure. Chinese models need to build layouts in both vertical capabilities and distribution channels.

Google has dropped a bombshell in the AI large-model arena. In the week 46 update of the Arena blind-test leaderboard, Gemini 4 Carbon debuted at No. 1 with 1,526 points, followed closely by its sibling Gemini 4 at 1,522. Google swept the top two spots — a "new model that immediately dominates the field" posture that is rarely seen in the arena's history.

Gemini 4 Carbon debuts at No. 1 with 1,526 points

Gemini 4 Carbon Takes the Top Spot: How Much the Blind-Test Score Is Worth

The Arena's blind-test mechanism ensures scores can't be gamed — users compare responses without being told which model produced them, and only then are votes counted toward the ranking. Gemini 4 Carbon scored 1,526 on its debut, 4 points above its sibling Gemini 4, showing that Google isn't relying on a single model breakthrough this time — the entire Gemini 4 series now has top-tier competitiveness. With two models occupying first and second place, Google has formed a "double insurance" policy in overall capability.

A new model that immediately dominates the field

From a competitive standpoint, this directly changes the previous situation where multiple models took turns at the top. Google has claimed two high positions with a single release, and the pressure on other vendors is real — they not only have to catch up with one Gemini 4 Carbon but also face a Gemini 4 at nearly the same level.

Technical Details of the Arena Blind-Test Mechanism

The Arena (typically referring to LMSYS Chatbot Arena) uses crowdsourced pairwise comparison: users are randomly given responses from two anonymous models and vote for the better one, after which the system converts the massive number of votes into comparable scores using an Elo or Bradley-Terry model. The key to this mechanism is the "blind" aspect — voters don't know the models' identities, so scores reflect genuine preferences rather than brand effects. But blind testing has limitations too: voters are mostly tech-community users whose preferences may differ from ordinary users; a single debut score is affected by sample size and opponent strength and has a certain fluctuation range. A debut score of 1,526 at the top means a statistically significant lead, but a 4-point gap between siblings is still close and requires multiple weeks of data to confirm stability.

Chinese Models Hold the Video Leaderboard: A Signal of Differentiated Breakthrough

Google took the top spot on the overall leaderboard, but Chinese models held their ground in the top-five group on the video leaderboard. This detail is worth a closer look: video generation and understanding is a direction with a relatively high technical threshold among multimodal capabilities, and the fact that Chinese models maintain a presence on this specialized leaderboard shows they haven't fallen behind in specific capability dimensions.

Chinese models hold the video leaderboard

The overall leaderboard measures the "versatility" of general conversation and reasoning, while the video leaderboard tests models' specialized abilities in temporal understanding and visual semantics. Choosing to establish a foothold in vertical tracks like video is a pragmatic strategy for Chinese models — rather than engaging in a head-on war of attrition with Google on general capabilities, it's better to build irreplaceability in specialized domains. Of course, whether the top-five position can be sustained will be the key observation point in next week's data.

Chinese models held steady in the top-five group on the video leaderboard this week

The Technical Threshold of Video Understanding and Generation

The video leaderboard evaluates models' ability to understand and generate dynamic visual content, involving specialized technologies such as temporal modeling, inter-frame consistency, action recognition, and visual-semantic alignment. Unlike text conversation, video tasks require models to maintain contextual coherence along the time dimension — for example, determining the order of events or tracking objects across frames. Current mainstream approaches typically combine visual encoders with temporal attention mechanisms, but the memory overhead and inference latency of long videos remain bottlenecks. The fact that Chinese models can stay in the top five on this leaderboard shows they have accumulated experience in video data training and temporal architectures, but the evaluation benchmarks for this track are not yet as mature as those for text leaderboards, and there may be a gap between leaderboard rankings and actual product experience.

Reddit Traffic Down 13%: AI Summaries Are Rewriting the Search Ecosystem

SimilarWeb data shows that Reddit's global traffic has fallen 13% over the past year, with traffic from Google Search noticeably decreasing. Reddit's response is to plan to require all visitors to log in — essentially tightening the traffic funnel and converting anonymous traffic into operable user assets.

The direct driver of the traffic decline is AI summaries. Google Search's AI Overviews provide answers directly on the results page, so users no longer need to click through to Reddit to read discussion threads. This is a structural shock to community platforms that rely on search referrals: content is still being used by AI, but traffic and ad revenue stay on the search side. Reddit's login requirement is also a way to find a way out of the dilemma of "content being summarized by AI while users don't visit."

AI Overviews and the Zero-Click Search Trend

AI Overviews are summary answers generated directly by Google Search at the top of the results page, allowing users to get information without clicking a link. This feature relies on large models to integrate search results, essentially compressing the "search-click-read" chain into "search-read answer." For content platforms, this means that even if content is cited by AI, traffic won't flow back. SimilarWeb's traffic data reflects exactly this structural change: search traffic has degraded from a "distribution funnel" to an "answer supply," and platforms have become training and retrieval material libraries for AI. Reddit's login requirement is an attempt to convert anonymous traffic into identifiable users to reduce dependence on search referrals, but it may also further reduce visits from new users.

The Next Step for Chinese Models: From Holding Ground to Breaking Through

Combining this week's leaderboard with Reddit's traffic changes, Chinese models face a trade-off between two directions. One is to continue deepening vertical capabilities such as video and hold their ground through differentiation; the other is to seek breakthroughs in general capabilities and strive to regain positions on the general leaderboard.

Reddit's experience also provides an additional warning: beyond model capability, distribution channels and traffic funnels are equally decisive for survival. If Chinese models want to become popular overseas, in addition to leaderboard scores, they also need to think about how to get users to actually use them and stay — rather than becoming the "anonymous content suppliers" behind AI summaries.

One-line Summary

Gemini 4 Carbon debuts at No. 1 on Arena; Chinese models hold the video leaderboard; Reddit traffic drops 13% due to AI summaries.

Paragraph Summary

The week 46 Arena blind-test leaderboard shows Google's Gemini 4 Carbon debuting at No. 1 with 1,526 points, followed closely by its sibling Gemini 4 at 1,522, with Google sweeping the top two and forming dual high-position competitiveness. Under the blind-test mechanism, scores reflect real user preferences, but debut scores still need multiple weeks of verification. Chinese models lost the top spot on the overall leaderboard but held firmly in the top five on the video leaderboard, reflecting a differentiation strategy. Meanwhile, SimilarWeb data shows Reddit's global traffic fell 13% over the year, as Google's AI Overviews provide answers directly and reduce search referrals, and Reddit plans to require visitors to log in to cope with changes in the traffic structure. Chinese models need to build layouts in both vertical capabilities and distribution channels.

#Gemini 4 Carbon#Arena leaderboard#AI model blind test#Chinese AI models#video generation leaderboard#Reddit traffic decline#AI Overviews#zero-click search#SimilarWeb#multimodal models
Share:

Related articles