Grok 4.6 Keeps Grok 4.5's Foundation and Improves the Training

xAI released Grok 4.6 on 7 August 2026. Like Z.ai's GLM-5.3 a week later, it reuses its predecessor's foundation - the same 1.5-trillion-parameter V9 base as Grok 4.5 - and takes its improvements from significantly upgraded supervised fine-tuning and reinforcement learning.

The focus is long-running agents and more ambitious interactive and visual work, while preserving the speed and token efficiency Grok 4.5 was known for. Reported positioning puts it against Moonshot's Kimi K3 and Anthropic's Claude Opus 4.8.

Coverage cites pricing at roughly half that of rival frontier models. On LMArena's text leaderboard Grok 4.6 has 1455 Elo (rank 72, high effort). A larger 2.1T-parameter Grok 4.7 is reported as scheduled a few weeks out.

Correction (2026-08-18): an earlier version of this item said xAI publishes no citable model documentation for Grok 4.6. That was wrong. xAI’s model docs carry full specifications and tiered pricing: a 500k context window, and $2 / $6 per million tokens under 200k tokens of prompt, doubling to $4 / $12 above it. Grok 4.6 is now on our roster with first-party sourcing. The 1.5T parameter count remains secondary - xAI does not publish a parameter figure.

Three vendors in two weeks attributed their gains to post-training: GLM-5.3, Grok 4.6 and DeepSeek V4-Pro GA all delivered their gains through post-training on an existing base rather than a fresh pre-training run.

Read xAI's announcement: x.ai