Musk admits Grok 4.8 trails Claude Opus 5.5; model still in RL after C/C++ stack rewrite

In an unusual public concession, Elon Musk said Grok 4.8 is behind Anthropic's Claude Opus 5.5, which he called 'a masterpiece'. He said Grok 4.8 is 'not as good' and predicted that Grok 5 could overtake Opus 5.5 within two to three months. Grok 4.8 has completed pretraining and is now in reinforcement learning. Musk is the only source for a reported 2.5-trillion-parameter size.

The delay has a technical cause. xAI replaced its training and serving stack, dropping Google's JAX framework and the SGLang inference engine in favor of a system written from scratch in C and C++. FourWeekMBA reports the model entered RL after a 57-day wait. Rewriting core infrastructure can give long-term speed and control, since a team can optimize directly for its own hardware cluster. It also carries significant short-term risk: new bugs, missing features, and months of engineering that does not directly improve the model.

The admission changes the competitive picture. Musk has typically claimed Grok leads or will lead. Conceding to Anthropic in the same week as Anthropic's IPO roadshow and the launch of Fable 5.2 gives Anthropic a credible endorsement from a rival. It also fits a broader pattern: OpenAI faces Reddit criticism of GPT-6 Instant, Google's Gemini 4 Carbon has latency complaints, and Meta has narrowed its focus to cost-efficient models. Meanwhile, xAI's recent shipped work has been incremental, such as retiring the legacy grok-imagine image endpoint and cutting image prices by a cent.

Skeptics point out that nothing about Grok 4.8 can be verified: no model card, no benchmarks, no pricing and no API listing. Musk's timelines have often slipped, and a 'two to three months' prediction for Grok 5 should be treated cautiously. Watch for a Grok 4.8 release with published evaluations, whether xAI documents the new C/C++ stack, and whether the RL phase delivers the improvement Musk is promising.