When xAI founder Musk answered a netizen’s question on X on November 10, “How does Grok 4.8 compare with Opus 5.5?”Reply says: "It's not that good. Opus 5.5 is a masterpiece." He went on to say that the next generation Grok 5 should surpass Anthropic's Claude Opus 5.5 "maybe within two to three months" and will continue to accelerate after that.

This Q&A took place after Musk announced the development progress of Grok 4.8. He didn't release test scores for Grok 4.8, nor did he give Grok 5 a firm launch date.

Grok 4.8 pre-training is completed, and will be given to Grok Bot users for trial this week

Musk said earlier on the same daypostZhong said that xAI has abandoned JAX and SGLang and rewritten the entire reinforcement learning and inference system from scratch in C/C++, which "took some more time." He said that pre-training has been completed, a lot of cleaning has been done on the supervised fine-tuning data, Grok 4.8 is currently in the reinforcement learning stage, and the evaluation scores have begun to rise.

He said he would first discuss the matter later this week GrokBot Users roll out Grok 4.8 on a small scale to see the reaction. He described Grok 4.8 as "nothing like 4.7 (thank God)" and "not amazing yet, but with the right foundation this time" and expected to evolve into "something special."

After netizens praised the switch to C language, Musk added that it was "C plus a little C++"; he also說, the new Grok training and inference system is derived from the results created by SpaceX engineers for Starlink. "We write our own header files and calculate our own time (including leap seconds)."

Opus 5.5 costs 40% less than its predecessor

according to Anthropic official introduction, Opus 5.5 is the first model of the Claude 5.5 series. Most work performance is equivalent to Fable 5.1. The cost is 40% lower than Opus 5 under normal workloads. The input and output costs per million tokens are US$4 and US$20 respectively, and the output speed is more than 30% faster.

In Terminal-Bench 4.0, which tests whether AI agents can complete complex tasks on the command line, Anthropic announced that Opus 5.5 scored 66.4%, and OpenAI itself announced that GPT-6 Astra scored 57.9%. xAI has not announced the results of Grok 4.8 in this type of test.

this article Musk calls Claude Opus 5.5 a masterpiece, but his new version of Grok cannot compare first appeared in Chain News ABMedia。