Grok 4.8 Has Not Been Released Yet—Musk Says He Doesn’t Care About Benchmarks

On October 30, Musk acknowledged that the development cycle for Grok 4.8 has taken longer than originally planned—but stressed that the model will live up to expectations. He noted that the team is not optimizing for benchmark test scores, nor is he personally interested in such metrics; instead, the focus remains squarely on the model’s real-world application value. As early as September 14, he publicly revealed that Grok 4.8 features 2.5 trillion parameters, is trained on a newly built, in-house C++ training system, and was expected to complete its pretraining phase that same week—followed immediately by reinforcement learning. Yet over six weeks have now passed since then, and the new model has yet to launch officially. According to the latest report from independent evaluation firm Artificial Analysis, under the unified xhigh inference tier, Grok 4.7’s Composite Intelligence Index improved by just 2 points over Grok 4.6—from 44 to 46—while its average tokens-per-question output decreased…