Grok 4.6 catches the frontier but loses what made it special

xAI acquired Cursor and used its RL/post-training techniques to improve Grok 4.5 into 4.6, now trading blows with GPT-5.6 Sonnet on benchmarks at a fifth the price. The model excels at long-running agentic tasks and multi-step orchestration, with real-world examples showing genuine capability in security audits and complex PR management.

The tradeoff is stark: Grok 4.6 loses the speed and token efficiency that made 4.5 uniquely useful. The model now generates 30% more tokens per task, making it more expensive and slower than its predecessor. What was a fast, cheap, reliable daily driver is now just another frontier model with comparable cost and latency.

Design and creative work remain weak points—Grok 4.6 produces noticeably worse UI than Sonnet or Fable, and failed entirely at a 3D game port where even cheaper models succeeded. For engineering leaders, the real signal is the rapid iteration pace: 4.7 is already mentioned as significantly better and just weeks away.

The practical implication is clear: Grok 4.6 is a transition release showing xAI can catch up technically, but the high-value niche of fast+cheap+capable is temporarily lost. Teams should watch cost-per-task carefully and evaluate whether the intelligence gains justify the regression in speed.