Grok 4.8 delayed; Musk says benchmark scores don't matter

Huo Xing Finance reports that on October 30, Musk acknowledged that the development of Grok 4.8 took longer than expected, but assured that the new model will not disappoint. He stated that the team is not optimizing for benchmark tests, and he personally does not care about scores, placing greater emphasis on the model’s practical applications. On September 14, Musk revealed that Grok 4.8 features 2.5 trillion parameters and uses a brand-new C++ training system, with pre-training expected to conclude that week, followed by reinforcement learning. More than six weeks later, the new model has still not been officially released. Independent evaluation agency Artificial Analysis found that, under the same xhigh inference tier, Grok 4.7’s overall intelligence score increased by only 2 points—from 44 to 46—compared to Grok 4.6, while the average output tokens per query rose from approximately 38,000 to 81,000. Users on the Cursor community have also complained that Grok 4.7 consumes credits faster without delivering noticeable improvements. During a CCTV interview in September, Musk admitted that Grok 4.7 falls short of Claude Opus 5.5 and estimated it would not catch up to the top tier until 2027. In October, he proposed allowing Grok Bot to invoke external models—including Claude—based on tasks, and plans to use the high-speed version of Grok 4.8 to handle large volumes of simple requests.