As Fable 5 starts slowly, Alibaba tests agents on real sales tasks
Fable 5’s first-month usage share remained limited as Tencent announced a larger Hy4 without a schedule, while Alibaba opened an arena scoring agents on cross-border sales tasks. New models increasingly need usage and task performance beyond announcements.
Artificial Intelligence··Evening
Fable 5 held a limited share in month one
Ramp’s August index puts Fable 5 at 6 percent of the tokens companies bought from Anthropic in its first month and 11.4 percent of what they spent there. On the OpenAI side GPT-5.6 Sol holds 25 percent of tokens and 23 percent of spending. Fable 5 lists at 10 dollars per million input tokens and 50 dollars per million output tokens. The index rests on payments that pass through Ramp’s own spend management product, so it measures corporate card and invoice flows rather than every deployment. The same data show Anthropic reaching 43.5 percent of companies in the United States in July against 39.7 percent for OpenAI, and a median artificial intelligence spend of 7,400 dollars per employee among the top 1 percent of spenders. The newest model is not automatically the most used. Fable 5’s spend share sits above its token share, suggesting costly outputs weigh more, yet both still trail GPT-5.6 Sol. Because the Ramp filter only covers payments through its own product, the figures represent the card-and-invoice slice rather than every corporate AI deployment. First-month data put concrete numbers on how interest in a frontier model can stall on price and habit.[1]
Hy3 usage jumps while Hy4 stays undated
Tencent said weekly use of its Hy3 model rose more than 68 times over the previous model after Hy3 left preview, and that it plans to release a larger-parameter Hy4 in the near term. The company disclosed neither a release date nor any technical specification for Hy4. Hy3 currently runs inside Tencent products including WorkBuddy, CodeBuddy, Yuanbao and ima. The same statement said WorkBuddy took more than 20 million PC visits in June. TechNode attributes the figures to GeekPark, which publishes in Chinese, and Tencent has released no model card or evaluation alongside the usage claim. The announcement ties growing use to a larger follow-on model, yet it gives readers no testable schedule or performance table. Because the 68-times jump is defined against the prior preview baseline, it does not by itself state absolute volume. WorkBuddy’s more than 20 million PC visits supply a separate product measure that the model is circulating inside shipped tools. Without a model card, outside comparison stays blocked. Set beside Fable 5’s measured first-month share, limited adoption figures stand next to an undated scale-up promise; both still ask for evidence beyond the announcement.[3]
An arena puts agents on real sales tasks
Alibaba Cloud has opened Qwen AI Arena, a challenge and evaluation platform where developers submit AI agent solutions against tasks drawn from real business scenarios. The platform also supplies entrants with the models, runtime environments and scoring tools. The first challenge asks entrants to produce product listings for the United States, South Korean and Brazilian markets, with English, Korean and Portuguese copy alongside product images and video. Automated testing is scheduled to begin in mid-August, after which the 30 highest-scoring submissions move to expert review. TechNode attributes the details to GeekPark, which publishes in Chinese. The arena seeks a third measure beside model announcements: a score on a cross-border sales task. Fable 5’s Ramp share tracks corporate adoption, Tencent’s Hy3 figures track in-product use, and Alibaba’s contest tests multilingual listings and media. New models therefore have to gather interest through measured usage and task performance, not only parameter counts or ship dates. Until scores appear, the automated round and the shortlist of 30 still ride on the calendar. Together the three tracks replace a press notice with concrete questions: who is using it, inside which product it circulates, and what score it earns on a sales task.[2], [1], [3]
