Grok 4.8 delayed, Musk says he doesn't care about benchmarks, only wants it to be truly useful.

Beating AI News Flash: Musk admitted on October 30 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 does not optimize for benchmarks, and that he personally does not care about scores, valuing the model's practical use cases more.

On September 14, Musk revealed that Grok 4.8 has 2.5 trillion parameters, adopts a brand-new C++ training system, and was expected to finish pre-training by the end of that week before entering reinforcement learning. Now, more than six weeks later, the new model has still not been officially released.

The reason Musk emphasizes not caring about scores is that Grok 4.7's results are truly not worth boasting about. Independent evaluation agency Artificial Analysis found that at the same xhigh reasoning level, Grok 4.7's comprehensive intelligence index improved by only 2 points over 4.6 (44→46), while the average output tokens per question increased from about 38,000 to 81,000. Users in the Cursor community have also complained that 4.7 consumes quota faster, while the actual improvement is not obvious.

In a September interview with CCTV, Musk already admitted that Grok 4.7 is inferior to Claude Opus 5.5, and is expected to have a chance of catching up with the top tier only in 2027. In October, he also proposed having Grok Bot call external models including Claude on a task-by-task basis, and planned to have the high-speed version of Grok 4.8 handle a large number of simple requests.