Tokenized Takes：大模型与AI流量跟踪器

Internet Research Analyst+1-212-250-9218 data trackers to better understand the state of play across frontier Al models and theimplications for our Internet coverage. In particular, we intend to track frontier modellaunches and the cadence between the releases, LMArena leaderboard standings, andconsumer engagement across the leading LLM platforms (ChatGPT, Gemini, Claude, MetaAl,CopilotandPerplexity),includingvisits,DAUs,sessioncountsperDAU,andaveragetimespent per day.We plan to then turn the lens to our coverage: LLM-referred traffic acrossseven verticals (in visits and as a % of total traffic), developer APl token share viaOpenRouter, what that implies for who is gaining or losing mind share, and which of ournames are best positioned as Al-driven discovery scales. Research Analyst+1-212-250-9357 Research Associate+1-212-250-2064 Research Associate+1-212-250-8153 On the evolution of the LLM landscape, our analysis suggests intensifying competitionacross both the enterprise Apl layer and the consumer engagement layer, alongsidewidening dispersion in model performance and monetization dynamics. ResearchAssociate+1-212-250-2060 Within the enterprise channel, OpenRouter data indicates a rapidly shifting leaderboard,31 May 2026),followed by Anthropic at 14.6%.Notably,Google-despite leading for muchof 2025 (peaking at ~45%)-has moderated to the #3 position (11.9%), while OpenAl sitsat #7 (~8%), behind three Chinese names Minimax, Xiaomi and Tencent. In aggregate,Chinese model providers collectively represent a substantial ~45% share. We viewdeveloper API spend as the most dynamic and competitive KPl in the ecosystem, with theenterprise usage mix diverging meaningfully from the consumer DAU landscape, whereChatGPT maintains a~4:1 lead. from its October 2025 peak of 6.2bn visits, while Gemini has scaled to ~2.9bn monthlyvisits, and Claude (950mn)is compounding most rapidly off a smaller base.In our view,the consumer Al market continues to consolidate towards at least a three-playerstructure rather than a winner-take-all outcome. From a model quality perspective, Anthropic currently holds a clear lead on the LMArenaText/Overall (6 of the Top 10) and Coding (8 of the Top 10) leaderboards, with Fable 5 beingthe current #1 across both.Overthe pastyear,leadership has rotated across three distinctregimes, with OpenAl ceding the top position to Google in July 2025 (after which Geminiretained leadership for ~7 months), before Anthropic assumed the lead in February 2026-highlighting the increasingly competitive and iterative nature of frontier modeldevelopment. We also observe meaningful divergence in APl pricing strategies. OpenAl's new flagshipGPT-5.5 (April 2026) lists at $5/$30 per million tokens - double GPT-5.4's $2.50/$15,though GPT-5.4 remains available as the cheaper production tier. The realized increase issmallerthantheheadlinedoubling:GPT-5.5completesthesameworkwithfewertokens,so the actual cost of a typical request is up ~49-92%, not the full 100%.Anthropic left its~35% more tokens for the same text, lifting effective per-request cost ~12-27% on a remain substantially cheaper -from ~3x cheaper (input,flagship-reasoning tier)to 15-100x (budget tier). We believe this widening dispersion underscores ongoingexperimentation with inference monetization and will likely determine which playershave thebalance-sheet flexibilityto sustain aggressive deployment strategies.Atthe infrastructurelayer, hyperscalers (notably GCP and AwS)remain well positioned to benefit from surging enterprise token demand on OpenRouter, which is running at >15xy/y growth. Meanwhile, both Google and Meta have placed flagship models within the Top10 of the Arena leaderboard. Meta's progress stands out: Muse Spark-the first releasefrom Meta Superintelligence Labs and Meta's first frontier-class model since Llama 4Maverick -debuted in early April at #3 on the LMArena text arena (tied with Gemini 3.1Pro and Claude Opus 4.6), against a #32 debut for the publicly released Llama 4 Mavericka year earlier. rapidly, reinforcing the early-stage nature of adoption curves. We do not however readthis growth as uniformly positive for destination sites: an assistant layer is increasinglyinserted between the consumer and the website.This could become a disintermediationrisk for names that previously benefited from direct traffic. On the other hand, the newtraffic LLMs create may also serve as an incremental customer acquisition channel,diversifying traffic originationbeyond just Google Search.In this report,we have analyzedthe impact of Al web traffic on different sectors. The underlying theme across everyvertical is a platform-driven inflection in LLM-referred traffic beginning February-March2026. We explore each of these verticals in more detail later in the "Al Traffic as a % ofTotal Traffic"section. Ads: LLM-referred traffic inflected sharply across Reddit, Pinterest and Facebookin April-May 2026, with May being an all-time high for each sit