Grok 4.5 Test: I Finally Didn't Disappointment This Time The first few generations of the xAI Grok were extremely popular before their release, but after actual testing, they couldn't handle it. So on July 9th, when Grok 4.5 was officially released, I didn't have much expectation. The result this time is a bit different, I lied to you that I am a puppy. I actually ran a coding task and the speed of Grok 4.5 is very fast, with programming ability firmly standing in the first tier. It also solves the problems that GLM 5.2 has not yet solved, and it is noticeably more streamlined. The reason why we have improved so much this time is due to a key information behind it, that the new model was trained together with the cursor. That is to say, it has obtained high-quality coding data and polished it in programming scenarios and high-quality data. It is priced at $2/M input and $6/M output, which is more cost-effective for overseas companies. Its price is higher than GLM5.2 and its ability is slightly stronger than GLM5.2. But it also has its shortcomings. The first weakness is the context. Currently, only 500000 tokens have been given, and the official says that the 1 million long context will not be updated until next week. You have to endure dealing with long documents and code repositories now. The second weakness is Chinese. Occasionally breaking links and making sentences less fluent when handling Chinese tasks, making it difficult to work with Chinese documents. Interestingly, at almost the same time, Tencent's hybrid hy3 was also brought up for discussion. The situation between the two is very similar: their abilities have clearly caught up, and the training data sources are highly similar, both from social platforms, accumulating a large amount of authentic and high-quality content. The exemption promotion of hy3 on OpenRouter is true, and the free call will continue until July 21st. You can also directly experience it on Tencent's own IMA, WorkBuddy, and Yuanbao. Tencent has such high-quality content sources as official account and Tencent News in its hands, and there is no lack of data base. HY3 itself is a medium-sized model, with 295B MoE, 21B activation parameters, and SWE bench running to 78. The scale is not large, but it performs outstandingly in Chinese language scenarios such as writing articles and documents. I have used it more than minimax-M3, ranking among the top tier in China. Putting these signals together, the conclusion is actually very clear. The upper limit of the model is determined not only by the parameters but also by the quality of the data. Claude has long been far ahead in programming, starting with high-quality interactive data from Cursor's early days. GLM was originally called Claude Pingti because it was the first to bet on coding and data accumulation from the beginning, so its programming ability has always been in the top tier of domestic production. Grok was able to catch up this time and trained together with the data from cursor. Tencent hy3 has both high-quality corpus from social platforms and unique user context from IMA and WorkBuddy. The combination of data and products provides a clear advantage in user experience. High quality data is the real reason why hy3 and Grok have caught up in this wave of model capabilities.