Key facts
- Musk said Grok 4.8 has finished pretraining and could begin a limited rollout to Bot users later in the week.
- Musk said xAI dropped JAX and SGLang for an inference and reinforcement-learning stack rewritten in C and C++.
- Musk said Grok 5 should exceed Opus 5.5 in two or three months, without giving scores or a firm release date.
- Anthropic reported Opus 5.5 scored 66.4% on Terminal-Bench 4.0 versus 57.9% for OpenAI's GPT-6 Astra.
- Musk has not published Grok 4.8 benchmark scores, so the size of the gap is not established.
Elon Musk said on November 10, 2026 that xAI's forthcoming Grok 4.8 would be "not as good" as Anthropic's Claude Opus 5.5, and called the rival model a "masterpiece" in a reply on X to a user named Dimdv.
The exchange began with a technical update. Musk said Grok 4.8 had finished pretraining but was still undergoing reinforcement learning, a phase in which a model's responses are refined using feedback and rewards. He said the team had spent considerable time cleaning up supervised fine-tuning data, the examples used to teach a model to follow instructions.
Musk said xAI had abandoned JAX, a machine-learning framework, and SGLang, software used to serve AI models, in favor of an inference and reinforcement-learning system rewritten in C and C++. Inference is the process of generating answers after a user sends a prompt. He provided no figures showing the resulting gains, and described Grok 4.8 as "not yet incredible" while saying its foundations were finally built correctly.
He said he expects a soft rollout to Bot users later in the week, giving xAI an early test before any broader release. He also credited Starlink engineers with work that helped produce the Grok training and inference stack, pointing to SpaceX's practice of writing its own headers and handling precise time calculations, including leap seconds.
Anthropic released Opus 5.5 on September 22, delivering performance comparable to its higher-end Fable 5.1 model on most work while costing 40% less to run than the previous Opus version on typical workloads, according to the company. On Terminal-Bench 4.0, which tests whether AI agents can complete complicated tasks in a command-line environment and scores them by the percentage successfully finished, Anthropic reported 66.4% for Opus 5.5 versus 57.9% for OpenAI's GPT-6 Astra.
Musk has not published Grok 4.8 scores that would establish the size of the gap. He said Grok 5, a next-generation model to be released after this update, "should exceed Opus 5.5," potentially in two or three months, without giving a firm release date.