xAI Launches Grok 4.6 With Improved Post-Training, Teases Grok 4.7

My take: The Grok 4.6 launch confirms a trend every major lab is following: at this stage of the race, the biggest gains no longer come from adding parameters, but from refining how the model learns from its own responses. xAI kept the same 1.5 trillion-parameter core from Grok 4.5 and invested in improved supervised fine-tuning (SFT) and reinforcement learning (RL). If it delivers, that means better results without raising inference costs.

For anyone using AI at work, this move matters because it signals that frontier model pricing could stay competitive even as quality rises. Access via SuperGrok costs $30 per month, and the API inherits Grok 4.5's pricing. Meanwhile, Grok 4.7, with 2.1 trillion parameters, is expected in the coming weeks.

The question worth asking is whether independent benchmarks will confirm what xAI claims. The company did not publish official scores on launch day, so performance figures circulating now come from the same lab that built the model. It is worth waiting for third-party evaluations before treating those numbers as definitive. What do you value more in a frontier model: more scale or better training?